The AI Vibe Check (Replay Vision)

Speaker 0: Thanks for coming out with the World Cup on and everything. I know there's some exciting games out shares. But hopefully, this is just as entertaining or or more. So this is AI Tinkerers. N/A, any 1st timers here? Speaker 0: Yeah. Oh, wow. A lot of 1st timers. Okay. Speaker 1: Cool. Speaker 0: Well, we're here to kinda show some demos, show some behind the scenes, hopefully Sean something from your colleagues. We do this once a month here in the Kin. The Kin is an incredible Code working space. Michael, you wanna give a little rundown of the space? Michael's the main man here. Speaker 1: How's it going? Yeah. Welcome to Kin. This is your 1st time here. We are a work club and membership club. Speaker 1: Wellness and stuff here. Just learning more about membership. Happy to speak to you about it after Speaker 2: seeing the the presentation. I'll be around. Cool. Speaker 0: They've got cool AI cafes that they do in the morning. Lots of AI Meetup. And upstairs AWS just Live a beautiful space if you haven't seen it yet. We also have PostHog here with us today. Game anybody use Posthog? Speaker 0: Do you guys know what this is? Yeah. We use Speaker 1: it all the time. Speaker 0: They've got some swag on the table over shares. Speaker 3: Some socks. Speaker 0: Is that what it is? Speaker 1: Socks, stickers, and keychains. Speaker 0: Awesome. Well, I'll let you give a proper introduction when you come up here, but we use their developer tools all the time. The session replays are great to watch how users use your product and, like, all the little analytics that they have. So we use it all the time. I am here with Hawxen Jun. Speaker 0: AI. We're a Meetup here in Venice, and we're a developer platform as well. You have access to all the different language models, all the video gen models, all the image gen models, the ability to fine tune them on your own data Dugan to store really big Date in, like, version really big datasets. So that's us, Oxen.ai. And then Sean over here, which you might know if you've been 2026 Tinkerers for a Speaker 1: AI, Speaker 0: got us all the egg rolls. Speaker 1: MoE Speaker 0: something called AI AI AWS is gonna be talking about the clock chain later automate. So that sounds super fun. We'll be doing 10 minute MoE. And when you're up 2026 demoing, just try to teach us something new. I will be sitting right here, and I'll just give you a little signal when you're at 3 minutes to try to keep things moving along. Speaker 0: Because last time, we didn't leave enough time for networking at the end. So that's what that's what our goal AWS. Today, We're gonna be doing some live demos. Demos might break. If they break, we all clap, and ask why it broke AWS AI we can learn from it. Speaker 0: And, yeah, have fun meeting new people. This is the agenda for today. I'm just doing the intro here, and then we'll have Kim come up. AI have Antoine come up afterwards. He has a really cool startup here as well that here he is. Speaker 0: Yeah. That does a lot of, training data for models, and they've, like, hooked into a game engine to capture data. And he's gonna talk all about the training data collection process and show us behind the scenes how they built this system. And then we Gargot Aaron here who is Wrote timer, but he submitted this project based knowledge Base in 1 easy templated directory, which I think about all the AI. Honestly, I'm like, where do I store all my all my context for automate m? Speaker 0: So I'm Replay curious what you've got cooking. And then Sean with the blockchain, which is just a really Wrote thing. Before we get in, who tried Claude Feywolf before it got taken away from us? Who built something crazy in the, like, 2 days that they had access? Yeah. Speaker 0: What MoE you go? MoE, basically, improved so much on the system we're currently building, and I easy, like, blown away by the Skills. And I let it, like, cook on a Friday. On a Saturday, I came in, and there was, like, the API error. I'm like, no. Speaker 0: I had the same thing. I woke up and I went to go send my next Speaker 1: message, and I was AI, Speaker 4: where did it go? But I'm not gonna AI. Like, GLM, 5.2 Oh. Which AWS, like, from a Chinese research team. Easy actually on top AI. Speaker 4: He's super close CTO. Speaker 0: Okay. That is funny use say that. Because I took a N/A shot of a tweet earlier automate. Speaker 1: AI if I Sean find it. I'm doing AI little screenshots. Speaker 0: But Base, it was AI Elon and people talking about when we're gonna get an open source. Mhmm. Elon was Live, early next year, no problem. And the z AI guys was Live, sooner June anything. Yeah. Speaker 0: And I used Speaker 4: to be honest, I think AI I subscribed yesterday Date 1 day I used AI video token. Speaker 1: That was live. Okay. Talk to him. Okay. Who else Speaker 0: built something wild Jun Live in the in the thing? Anyway yeah. Speaker 5: A little more boring. We had a huge customer that had a nasty race condition on session replay. It had been Date, you know, for, like, 6 weeks. Big customer. We don't have an answer for them. Speaker 5: All of our agents were on it for weeks Jun Base in about 30 minutes solved it. It's a crazy ass race condition that none of us could figure out. So Speaker 0: We solved the race condition teaching too. We're capable of guessing. Crazy. And And it was between 2 ours particularly easy between 2 different code bases that Wrote talking to each other. So, I gave it both code AI, and then I put up a PR for both. Speaker 0: And I was like, Speaker 3: that fixed it. Speaker 1: Cool. Speaker 0: Well, I I think with Claude, Fable coming out, it's AI you're kind of at the point where you can just build almost anything that comes into your mind. So I'm constantly thinking, like, are you building the right thing? Because what I spent my Claude table tokens on was I didn't wanna learn a video editor Software. So I just, like, by coded a video editor for AI, and it was fun. It was, like, I've spent half of the Saturday doing City, but then at the end of the day, I was like, why didn't I just spend that time learning context? Speaker 0: Like, I don't know. So just think about that when you're using these things. It's like, are you building the right thing? AI really enjoyed this YouTube video from Jeremy Howard on this topic. If you guys are curious, he talks a lot about I mean, if you don't know Jeremy, he's AI 1 of the best educators in the space if you wanna learn about AI. Speaker 0: But he was talking about how these models, have this really weird property of use can really easily get out of flow. Like, I don't know about you guys, but my ADHD just go all over the place when I'm using these things. And there's this nice chart here where you have skill on the y axis and challenge on the x axis. And flow is, like, right when you're at the edge of your skill level Jun it's challenging Jun you're in the Los. It's AI when you're playing tennis against somebody, you wanna play tennis who's, like, somebody that's slightly better than you. Speaker 0: That's, like, when you have the best tennis match. You don't wanna play that person who's just, like, way down here, and you don't wanna play Roger Federer. So I feel like it's a similar thing with AI. And, like, if it's too challenging and your Skills too low, you get anxiety. And if it's not challenging AI and your Skills Replay high, you get bored. Speaker 0: So we're aiming for this, like, upper left category. And I feel like sometimes AI can just be Live this slot machine of building the wrong thing because you get these dopamine hits every time. Like, I just build this timeline every day. You get these dopamine hits every time. Like, I just build this AI every day. Speaker 2: That's so cool. Speaker 0: Why? Like, you're Speaker 3: just gonna see that. Could you go back 2 pages from here just to get it? This 1? Yeah. Speaker 1: The other 1? Just like it. Oh, yeah. Yeah. Speaker 0: API video right here, highly recommend. On-demand so anyways, like what is a right teaching? And my takeaways from this video was build tools that kind of help you on your on the 3 these 3 main teaching. Your autonomy, your mastery, and your Opus. AI, if you've heard those 3 things before. Speaker 0: So build things that, like, help you in your thinking process, not take it away from use. Help you grow MoE you're, like, playing that tennis player that's, like, a little better than you and you're Hooking. Or something that gives you Opus, AI, helping other people. Sometimes the purpose can just Vibe, like, building something cool, but just try to think about those things Jun your vibe Code, and let's not just waste our tokens on random shit. That's all. Speaker 0: Let's hear from PostHog. I'm excited to see what we got. Speaker 5: Thanks MoE the plug as a replay user. That is my team. Speaker 1: I Speaker 5: love replay. Hopefully, I love this. So hi, everyone. I'm Kim. I work at PostHog. Speaker 5: I'm a product Engineer, and I'm talking about what I am calling the AI vibe check, which is analyzing Jun use experience at scale and figuring out what to build from that, using session replay under the hood. So Speaker 6: got a great intro already, but just really quick Speaker 5: PostHog talk, AI of the all in 1, product analytics, deploying type products all in 1. This slide is lead out of date. We've added new products, and we're kind of going into the more self driving product direction, which is also what we talk really about today. I forgot that. How many of us actually know if our users are having a good time while they're using our products? Speaker 5: We have all these great analytics tools, but they're kind of the shadow Plato's cave. AI, they're telling us what happened robust not why. Speaker 1: Anything Speaker 5: about the qualitative experience in the middle unless you remember to track it, which a lot of these things are more AI based and you can't really track anyway. For example, you might actually get a successful user conversion but you might be missing all of this context underneath the hood of the terrible time your user had getting there and missing valuable feedback that you could Vibe, incorporating HINTS Speaker 1: your feature. Speaker 5: So So a SSH, session Replay, my team, fantastic. You can watch what your user is doing on Hotflix which is a great not Netflix Wrote. And that should be the end of it. Right? Like N/A we can literally watch what our users are doing. Speaker 5: We can see when they hit friction. We can see when they HINTS, dead ends or broken HINTS, things we're not capturing in Event. And that should be it. Perfect. Except this introduces an entirely new problem which is the problem of scale. Speaker 5: So teams are having thousands of replays come in. They're great data Live, but they're really expensive to watch. It's time consuming. Aaron my 1st week at PostHog hog, we had teams telling us about how they would have watch parties. They would try to assign a replay to every person at their Community MoE everything would get watched. Speaker 5: And it's just mind numbing. And also if you're 1 human, you're not gonna be able to synthesize what you saw from replay 1 10 to 50. Like you're you're gonna lose all of these, patterns and trends underneath the hood. So we turned to AI to solve the solution or to solve this problem and we've kind of had a winding path to get there. So middle Flash last year, a different team started trying to build event based summaries of sessions. Speaker 5: They were really janky. They looked great robust if use, like, squinted at all and you AI it was, like, half hallucination AWS it was still relying on that, like, Plato's shadow in the cave type data. And then Deploying of this year, we started looking at doing r r web based summaries. And to step back, r r web is the giant giant JSON file that powers, session replay under the hood. It's tracking every single DOM mutation, every single mouse click. Speaker 5: It's a huge file Jun it actually works. Like the LLM can parse it robust they're not actually trained on RRweb So it just blows up the token context. It's really slow Jun it's not gonna work at scale for our users. And so we've come to our final not final. Our for N/A, we're shipping City Session, which is video based summaries. Speaker 5: So we are taking these Wrote snapshot replays, rasterizing them into video at 16 x speed, cutting out all of the inactivity periods Dugan feeding that to Gemini. And we're actually getting it sounds crazy, robust, like, it's cheaper than r o web and we're getting really great results, like amazing results. And next AI going to show a demo of that. So I guess to lead track, we're calling it Vision Jun even though the peep the thing people were asking for most was summaries, we kind of were Live, why stop there? We're allowing people to configure any output. Speaker 5: So not only can you ask for a summary from the agent, you can ask for, like, Base. Score, you know, user happiness on a scale of 1 to 10 Wrote tag it with 1 of these 3 session types. So you can configure all these different outputs Jun it will AI would after you've user AI, it'll, you know, give you output based on that. And then let me just check. Actually, we're just gonna go for it. Speaker 5: I prerecorded. I'm so sorry. I didn't know what the Internet situation would be. And like I said, this is just the summary SSH, so that's AI of the most asked for. Actually, MoE sorry. Speaker 5: This is happening so live. I'm just gonna rewind. If you can't tell, this is my 1st time using this. Speaker 1: Did you drive kind of video editor? No. No. Okay. Speaker 6: So the team wants you to watch every single replay that comes in every week, but it's just getting to be way too much. I don't have time to scroll through all of these. So let's see what something like Replay vision can do for me. Let's create a scanner. We have some templated. Speaker 6: That's helpful. Session summary seems like easy Esther go. We can keep all of the defaults here, go to triggers. We're still a relatively small Meetup, so I think keeping it at 100 is probably good. But we can start, you know, sampling as we get Builder. Speaker 6: Maybe add some filters CTO, like, certain page views or something if we wanted to focus on a certain flow. But for now, my PM wants everything, so we're just gonna create a scanner that scans everything. Speaker 1: Cool. And then no Speaker 6: replays have come in since this schedule was set up 3 seconds ago. So let's see what we can do with some on demand, scanning. So let's just fire these off. Speaker 5: The UI is another thing. It's the Speaker 1: new product. A large selection of these. Speaker 6: And AI we can go back CTO the SSH. We see they're running. Some of them are ineligible. That's actually awesome because it is saving me per the quota by not wasting a observation on Angeles recording. But we can see some are coming in. Speaker 6: So again, even this, let's say, I don't have time to go in and dive into these. Let's just ask Los AI AI Live, like, hey. Find a group summary for MoE. Find common themes across these recordings. In the meantime, while I'm waiting for that, I can take a few moments to look at myself. Speaker 6: Oh, no. Log in file in the title here. That sounds great. So we can see invalid username or passwords errors, significant frustration, rage Claude. See the implement here. Speaker 6: Yeah. Jun doesn't seem super great. And let's see what our group summary found. Search is broken. Cool. Speaker 6: Very important to know rage clicks and text clicks that kind of echo what you just saw. Speaker 1: AI up drop off. Awesome. So from this, I can go back to my PM, Speaker 6: send them this exact report teaching them kind of what the Replay from this week found, and then we have a super clear priority list of what we need to work on or fix automate this a newer experience for the users going forward. Cool. Speaker 5: So that's the demo for that section. And like I said, you CDN configure different kinds of output. That's just summary. And I just did on-demand, but you can set it up on a schedule. It file run every 5 minutes on everything that's coming in so we get AI updates. Speaker 5: Learnings from this journey so far. For prompting for us was kind of a bit of the art and science. Our model was way too nice. Like when we would try to test out our scoring, for example, like how successful was the user in their talk, we would always get 8 and above even when we AI purposely failed. So we'd have to be AI, Code. Speaker 5: No. When we say 0 to 10, Live, we mean use the full spectrum of 0 to 10. And on that note, it will try to find meaning where there is none. So that's we kind of arbitrarily cut off at 10 seconds, but we might tweak that. But if we would feed it a 2 2nd video, it would extrapolate AI meaning from that. Speaker 5: And so just throw those away. Kind of obvious business context is super helpful. We've experimented with kind of adding almost like an Jun to PostHog MoE you can drop your internal docs, you can drop a lot AI easy kind of business context that the agent can pull from to learn more about your business Event though it's pretty good at just extracting it from the video. And then the final learning, which is kind of Live what I'm excited about going forward, is that this is automate, but I feel like the Jun then what else next is kind of where the magic is gonna happen. MoE, for example, Live, that was me manually going through the summaries and sending it to my PM. Speaker 5: But, like, if, the scanner AI, like, hey. You have dead clicks on search. Like, it should just be able to put up a PR for that. And so hooking that into the life cycle of product, I Tinkerers where because it's not just gonna happen, maybe alerts, you know, schedules of, like, hey. Send me a weekly digest of what's happening. Speaker 5: If that rolling average drops below 5, you know, let us know. Things like that. Yeah. This launched, like, 3 days ago. It's in Date, so it's definitely still rough edges. Speaker 5: We're learning a lot. We're putting the tracks before the train a little City. Robust, yeah, any questions for 1 minute? Sure. What's the hell Speaker 1: on with that? It's Speaker 5: Gemini. Gemini Flash. That was like Speaker 0: so nice. Yeah. What did you have to say to it? Speaker 5: I think that was the person MoE did the multiple iterations on the score, it AI it was like, use know, you want to be easy to a little bit. It's file, AI. Like, use the Wrote scale. Like, it is it is okay to be a low score. It's not bad to be a low score. Speaker 5: You're not being mean to me. I mean, you're a low score. Speaker 7: Yeah. Because, maybe you was it synthetic users? Speaker 5: What do you mean? Speaker 7: Like, could you have an AI pretend to be a type of user? Speaker 5: As, like, a tester, you mean? Speaker 7: Or Yeah. Yeah. I well, I guess I'd be curious if you could take the data you're building up and train synthetic users. Speaker 5: Oh, interesting. I I do want Posthog Speaker 7: No. Sorry. No. No. As a developer, you know what I want Posthog to give me? Speaker 7: Yeah. It's an endpoint where I have a beta demo. Speaker 1: Mhmm. Speaker 7: And I run it through there and it says, here's 12 synthetic users. Totally. There's an open source project called synthetic users. Yeah. But I I can't help but think that it would be the next No. Speaker 7: For sure. To do with this is, like, pretend to be those users. Speaker 5: Yeah. I think it's come up in hackathon. People have tried to like, they come up for sure. But it doesn't exist right now. Speaker 4: You have the data? Speaker 5: We exactly. Like, literally. Yeah. Just plug in, like, a playwright m c p and then just build. Deepens for you? Speaker 5: Yeah. How do you choose how and why did you Speaker 1: choose Gemini? Speaker 5: That was on my decision. I don't actually know. Can you give me a video? Speaker 1: Yeah. I'm not sure. It's best with video. Yeah. I think that's that's yeah. Speaker 1: How did you Speaker 5: turn it into video? Speaker 1: This, Speaker 5: that I also can get back to you on. Yeah. A very a cron job that's running a lot of random stuff under the hood. Or temporal, not cron, temporal. Speaker 1: So especially when it comes to, like, emotional state of the users to be correlated Speaker 5: Oh, yeah. Speaker 7: We have wearables Speaker 1: in there. Speaker 5: Interesting. Do you have, like, heartbeat data? Speaker 1: It's nice to aura ring. Oh, nice. You know? Speaker 5: Interesting. Let me know if you wanna test out vision. Speaker 0: Did you see that tweet with the guy who was, like, I I looked at my, I don't know if it's aura or whoop for every call I had with my Speaker 1: AI Speaker 5: we've called that AI, like accurate mode Los, like a 16 by 16 grid or something shares it's all similar paths. Speaker 1: Super amazing. Yeah. Speaker 5: It's it's come up a few times Dugan in various hackathons. We've AI to build that Jun it's never quite worked out that well for us AWS every Base soon as it gets offset by a little bit, you're kind of AI losing City. But it it's come up enough where I think there is, like, something there. We just haven't done that yet. Yeah. Speaker 5: True. We are. That is kind of the, like, the thing we're figuring out right now is, like, pricing. We listen. We Jun joined this on 3 days ago to a very small Los beta. Speaker 5: We have no idea how the costs are gonna explode. Like, we have some training wheels and guardrails, but we're really trying 2026, yeah, deal with the question. I don't wanna Speaker 0: That's great. Got it. Thank you. My favorite posthog feature or Easter egg was when I saw the rage click, event. Speaker 5: Oh, yeah. Speaker 0: You guys CTO. Speaker 5: Yeah. Speaker 0: It's just like You have Speaker 5: a do you have a hedgehog voucher now? Speaker 4: I'm not sure. Speaker 5: You should have hedgehog voucher. Okay. Speaker 0: You're gonna have to show me. Okay. Speaker 1: I see. Speaker 4: Alright. Well, I wanted to share a little bit about my journey. I Event, like, 6 months, building a company, building the whole stack on my own, and I AI wanna share about what I discover and who Vibe been using AI actually, so you guys can, like, replicate that. So Base AI I've been, like, working towards is Skills sandboxes and fast feedback, which I understood were, like, the key component for, like, an AI to be good at its work. AI. Speaker 4: So what we do at Originlab is Wrote turn game engine into AI training data. So the way we do it is Base we kinda hack all the possible games and game Engineer in the market, and we reverse engineer graphic engine and game engine to be able to extract Date normal action and Event, but also camera poses and much more things from the graphic engine and the game engine. So the way we can do that is actually by using a lot of AI to reverse engineer a lot of those games. And then the idea is Live we sit in the middle, so we help game studios to make an additional stream of revenue. As you guys know, AI, game industry is not doing AI great these days. Speaker 4: So what we're doing is basically we help them, AI, like, monetize their information to sell it to AI companies. And so we Building all the different modalities, Game sure that it's the best quality to sell it to, like, AI companies. So for robotic, for Folder, and asset generation, but also for gonna show you quickly our platform. Hopefully that will okay. So this is basically a showcase of what we do. Speaker 4: So basically, the idea is that we go to any game Jun we'll extract a bunch of different AI modalities. So here, there's a player being in, like, a specific world that AI, like, RPG. What we'll do is, like, we remove the HUDs AWS the HUDs are, like, in-depth, like, artifact for, like, AI training and stuff. So, like, all those companies, they don't wanna see that. They wanna see, like, Claude kind of, like, SSH, understanding SSH, and things like that. Speaker 4: And so what we do as well, I don't know if you see AI right here, is Base we trace the camera within this world. So we reverse engineer the game engine to be able to trace that. In addition, we grab the keyboard and the mouse action, but also the depth map. Like, this sample doesn't have depth map, but I AI show you 1 with depth map. And the idea is, like, we are able to track everything within a specific Game. Speaker 4: So this is, like, perfect for robotic. The idea is, like, we wanna have the most diverse Date as possible on the world to be able to pretrain a lot of different AI. So we do a bunch of different things. We do, like, basically FPS, we do adventure game, and we help video studios actually license their game. So AI company buy the license as well when they buy AI product. Speaker 4: I'm gonna show you as Self, basically, what we do. This is AI a demo of our platform. So this is we have 3 Product. We have a system that basically mod games. We have a system that does recordings. Speaker 4: MoE, like, that's a recorder. Think about, like, the Opus, but, like, the superpower OBS that does a bunch of the different things. And then we also have a platform where people can actually actually make money by uploading content and playing games, but also people can actually buy content from us. The idea is we come in the middle and we ensure that the data is clear and good for everybody. And so AWS you can see here, basically Simple is playing your game and we actually get the depth map. Speaker 4: I don't I don't know if you see it well, but, basically, you get the depth map in a N/A Replay time. This is a really complex product because you're touching Date, like, classic engine, and it's AI mystery world for, like, most of the people in the world, I Skills say. And it's Live only a few people know that. And so we're able to do that using a lot of AI, I'm not gonna lie, and using a lot of, like, attendees recognition because it's really good at that. And so I'm gonna go alright. Speaker 4: Alright. MoE, basically, from my experience using AI, and I think I reflect a lot about what you said, AWS, like, I I felt like I wasn't in the flow. Like, I was, like, building stuff but not scalable. And but with this challenge, it was, like, really, really hard Knowledge. It's 1 of the hardest challenge in technology, Jun so that's why it was super, super exciting to work with. Speaker 4: And so, basically, I think, like, what AI is really bad at is produces, like, Jun Eloy. I Code Date, it it's fast to write, but it's painful to maintain. We build System that don't Folder up. Like, use worked today, but then after, like, it's really brittle to to work on the long term. And you always kinda, like it always implement. Speaker 4: It doesn't, like, creep it doesn't really, like, focus on, like, the core component of your application. And so I think the biggest end blocker for AI is r and d. So through pattern recognition, it's really good at reverse engineering, like memory pattern, understanding, like, basically teaching able to read the memory game, stuff that I AI would never be able to to have done before. And then, we also really fast at prototyping. So, like, what will take, like, 2 weeks for research now takes 1 day to validate, and then we can just, like, proceed into, like, getting a new modality, being able to add that new stuff. Speaker 4: Also, yeah, the research has Skills. Like, I use Meetup research every day. Like, every time I have an idea, I will type Jun, and then I would create a report and AI organizer my my idea like that. I think, like, deep research is really, really cool tool to have idea implement, Jun specifically when it's, like, super Date AI, like, explaining. When you talk about, like, c plus plus and super, like, low level programming, this is really useful. Speaker 4: By the way, I Event do any, like, game programming. I didn't do any, like, reverse engineering in the past. I learned everything in the past 6 months. So the idea is, like, we give every agent a skill and a sandbox. Speaker 0: It's Speaker 4: what I've been really good is not actually going straight to the problem and the solution with my agent. I will build all the different tools, all the different System that will help my agent get better. The idea is Live CTO define tools, also structure languages. I never done Rust. I started to do Rust with AI because it's a really verbose language that whenever you compile, will tell you once things are wrong. Speaker 4: Python is really good at telling you yes and then breaks in the middle. Also a lot of different gutters as well Jun a lot of AWS. Like, log as much as possible because AI is really, really good at going through logs and understanding problem by isolation. So the goal after doing all of that is basically you fail fast, you get the most context as possible, you iterate, and you fix. And so it's really like a step by step process that you have to follow. Speaker 4: So the Speaker 1: result is now I'm Speaker 4: able to mod and reverse engineer a game that would take me 2 weeks into 15 to 30 minutes, for Unreal Engine 4, 5, DX 12, DX 11. And so the way I do it is Base Sean AI is controlling a bunch of different Windows machine. It's a Claude code or GLM 5.2 AWS I started yesterday. That is basically controlling a bunch of machine through SSH based starting the game, reading the memory of the game, implementing MoE, and starting hot reloading the mod as the game is still running to be able to scan all the memory, getting everything, and diagnose also the rendering pipeline AWS those are, like, really hard to see. I mean, that's where we get, like, depth map, normals, and this type of things. Speaker 4: Jun, and then afterwards, like, the MoE is actually distributing the mod into our application. So next time a player comes in, they just have to say, I wanna play this game, and then it auto talk our mod, and then they get rewards by playing the game. All in 1 SDK. Basically, the idea is, like, as I told you, building the Los so AI can be really efficient at it. We're basically spending a lot of time building this a SDK, and the AI is only, like, filling the gap. Speaker 4: Right now, SDK Base, like, 1 game to be modded. It's just, like, 20 lines of code to mod Jun game. Alright. So, actually, I Date this Project, and then I implement for, like, the other people in my team. So through the design, the product scoping, the data model design, the rest API design, the coding, and the auditing, everything is Skills based going through each team and being audited by each team. Speaker 4: The MoE specialized and driven the agent, the more reliable it become. I will never start a session and be like, I wanna do that. I will always do deep research, try to see how it can window of scope and design Esther, and then MoE into the implementation, and I will always do end to end testing. The the AI is so good at running a to a end to end testing and being able to debug from that. So, like, the biggest learning, human on the architecture. Speaker 4: Never let the AI use your architecture. Otherwise, like, your project will be doomed within, like, a few months. Bit file back loops for the AI. So try to build your System so the AI learns as fast as as possible. And lean on remote execution. Speaker 4: So give, like, a bunch of tools for the AI to call different things. Like, I really like skills. I'm not a huge fan of MCPs because it has had a lot of noise, but you can build your tools. Like, it's a simple, like, bash command that you can build and AI Code, actually. And so I have, like, 20 Windows SSH with different GPUs that are running simultaneously. Speaker 4: And the AI is, like, Kim, like, configuring each of those. And AI didn't Replays engineering. I don't wanna, like, say that because I don't think it's true. I think it amplifies the workflow, and engineers spend much more time, like, building a well designed system. So building the right tool, having clear process, and being able to build AI fast feedback. Speaker 4: And, yeah, that's it. Okay. Speaker 2: If you think about Kim you think about taking Steam Proton, which is open source and instrumenting it MoE you can get directly right in talk the calls as they happen. Speaker 4: No. We did an Steam Proton? Speaker 5: Steam Proton. Speaker 4: Steam Proton. So it's like a Jira on top of, it's an Speaker 2: advanced AI Uh-huh. SSH. So Windows emulator for Game. And it runs the majority of Game, not not every Game. But the based majority of games, it runs really well. Speaker 4: Oh, okay. That makes sense. Yeah. So the thing is, like, actually, everything that we do is on the lowest level. So the extraction, the enabling, all Avenue on the GPU. Speaker 4: So it never leaves the GPU. So as soon as we use, like, another layer that is on top of it, then we lose the ability to So you Speaker 2: actually need to go directly to the AI? Speaker 4: Yeah. Speaker 2: So Okay. Speaker 4: We use, like, NVIDIA deliver. So as soon as NVIDIA does an update, it might break some of our Esther. But we try to window, like, be N/A careful about that. Speaker 1: With Speaker 2: the whole Steam SSH and Steam Deck, that might not work directly because it goes through Linux. Speaker 4: Yeah. So it's it's really different track to our Linux SSH. And it's all about what we decide to port as kind of Live operating system and drivers as well. Speaker 1: Yeah. Speaker 4: I Kim of feel like there's a lot of Windows player out there and a lot of Yeah. Ability to yeah. But I do I do agree with that. And for the longevity, we wanna do like metal, OpenGL, and also, like, graphic dropper as well. Speaker 1: Any other questions? Who's using the data Jun how are they using it for pretraining? Speaker 4: Yeah. MoE, basically, they use it to pretrain robotics. So AWS we're able to, like, capture, like, hours at scale, like, potentially by the end of the year, we'll N/A have, like, 1000000 hours of content with, like, depth map local and everything. And so it helped them, Wrote train robotics to understand, like, kind of Kim Hooking stuff together, understanding interaction with AI and stuff. It's not perfect AWS, of course, it's not AI real world Date, but it's good enough to Wrote train a model to understand, like, causation, effect, how do you interact with an environment, those type of things. Speaker 4: And also it's so diverse and so creative that you have so many Code ideas that you accurate explore the the MoE. Like, you can, I don't know, crash into a Claude? If you do, like, robotic teaching, Date, like, there's no way you can keep, like, crashing into a car that Eloy still have to do, like, SSH, stuff like that. Speaker 1: Yeah. Speaker 4: Code. We don't do that from VerQuint. We do that from from the game engine. And so the idea is, like, we give the AI, a bunch of tool to scan the memory on its own and give, like, the AI just have to say, oh, this address, this pattern, there is the camera, and then we can actually, like, keep the pattern even on updates and stuff. And so we read from the game engine. Speaker 4: So we do support Unreal June, 4 and AI. We do City, and we wanna do also proprietary engine. So we give our system CTO, game developer, and they just have to give us, like, memory address on a teaching we wanna capture, and that's it. Speaker 1: I know there's tools like that map and the other sort of thing. I know there are those tools. I could are you using those or just reentering? Speaker 4: No. So we N/A, like, actually use AI for most of those AWS those are, like, kind of Live the licensing on those tools. It's kinda AI you're Speaker 1: not Speaker 4: SSH% sure how much you can use those tools Jun all. So I I kinda wanted myself to go on a challenge and actually learn on the Use, but I get super specialized to what I want to do. But, yeah, there's a bunch of 2026 that we are, like, kinda, like, maintaining and, like, expanding. Speaker 1: Yeah. No. Speaker 4: Yeah. It's basically reading memory of a game. It it doesn't know anything. Like, for Unreal Engine, it knows, like, the API because it's Yeah. It's available. Speaker 4: But for some of those, like, it looks lead how the how the camera MoE. So, like, the transition in your view Jun then do, like, matrices of, okay, where in the memory actually the value use AI to the camera movement. So it's insane what AI is able to do. Like, I would have never been able to do something like that with my AI. Speaker 1: Oh, yeah. Speaker 4: 100%. AI fighting with easy and tight SSH right now. I think it's because they're AI they keep, like, flagging us all the time. So AI able to bypass it, but not on the multiplayer teaching. Yes. Speaker 1: Maybe 1 day. Speaker 4: Alright. Thank you. Speaker 1: So Speaker 2: I have no slides whatsoever. I love that. Yeah. And I have a framework. Yes. Speaker 1: MoE I have no slides whatsoever. I love that. Yeah. Speaker 2: And I have a framework. Speaker 1: Yes. Speaker 2: Jun I suppose you can say this is a demo, but it's also There we go. Okay. MoE, it's going to Speaker 1: there Speaker 2: we go. So let's, unify. There we go. So briefly, as some of you have found, knowledge Base, information is, really implement, and how your agent can find information to quickly get to start working can be a challenge. So Replay, I'll start actually with the template. Speaker 2: So my entire knowledge base is just this directory on my machine. Each 1 has, Base, directory point. So a document map and a project brief. What is this? Right now, obviously, this is, fairly empty. Speaker 2: I'll show you an actual 1. But what is this? So in AI, I can go, okay. Look at this and figure out what is am I gonna start working on? It's not necessarily ready to start working, but maybe that's enough for it to know, oh, I'm looking at this type of Project. Speaker 2: So AWS the code is over here on this other directory. Let me go start looking there based on whatever you prompted it for. In addition, the document map is incredibly helpful for durable artifacts. So your deep dive learnings, your deep dive research references, APIs Origin whatnot that you happen to have it go research, or your current, tracker of what's my plan for implementing this 1 Capture, or I'm reviewing this PR that I have lots of back and forth with. So some of them are are meant to be durable. Speaker 2: Some of them are, AI, Check, kind of a general whatever view stuff that's useful right now, but then it'll go away and eventually clean it up. Kind of a June, historical artifacts that you want to keep around that you might need to kind of what did I do in the past? Output, various outputs that you might need to take out of the the knowledge base to get elsewhere. Overall status, decisions you've made. How many times have you made a decision and then you forget to store it somewhere to surface it back Jun. Speaker 2: And then you end up reevaluating and remaking that decisions. AI, the backlog, so I have a quick idea or, oh, hey. I'm in the middle of something. We found this thing. Let me put it somewhere that I can come back to it because I wanna continue fixing or doing the task I'm on, but I need somewhere to store it that is obvious to get it back around to. Speaker 2: Now on a specific task, it might be it, an own task, backlog or whatnot, but not necessarily. It all depends on your structure. And then, obviously, what you always start loading, the project brief, what are we working on, what type of thing is this. That can include also what tools, languages, or other stuff that needs to be aware of. So this is, obviously, kinda useful for us humans, but really useful AI found for for my agent. Speaker 2: Jun plans is ongoing plans. And these Readmes, is partly for humans, partly for, Jun. So the basic introduction on the lead MoE is for us Gemini, what am I trying to solve? How do I configure my Jun? So I just say, here is in in your system prompt, here is where on my file system the my knowledge base is. Speaker 2: Each 1 is a sub-directory, and then that's it. You just create that directory. You put this template directory in there, and then you create your subdirectory projects MoE tell the agent Code Date a new project based on the template. Here's the name of it. And down here AWS, here is how agents should use this. Speaker 2: If you're an Jun reading this, here's the protocol. So you don't actually need to do anything. You tell it Date a new project. It prepopulates by copying in all these base files. It prevent up your project brief as you work with it to set up what this new project is, and all that sort of stuff. Speaker 2: AWS, things like write back. Make sure that you write back regularly to to teaching, especially when you're dealing with tracking files, plans, where am I in in a particular task. So that's, that's kind of generally the Meetup that I have. So looking at something that's kind of real, a idea that I'm iterating on that AI not started implementing yet, which is actually to replace this, with accurate a a full text search, vector search, information retrieval system. So turn this into more of a PostgreSQL track AWS if this is gonna be a knowledge base for all time, eventually, it's gonna get to be too AI. Speaker 2: Like, having Claude Event though it's different directories, if I wanna know sort through all these things or I have a bunch of 1 off little notes of AI and thoughts, like, I have a Pebble Time 2 coming. Little AI. I push a button. I speak into it. There's a 10 2nd note. Speaker 2: Okay. But how do I surface that later? And so my idea is it'll end up coming to my agent, through some kind of, queue, System, And it will read and go, okay. I need to add this to the personal. I need to add this to this work project, whatever. Speaker 2: Adds it to the index, categorizes what types of things this is, tags it, and then puts it into a whole Track category system, so that I can easily MoE, okay. So what were my ideas from last week? I had this idea of something like this 2 weeks ago, I think it was. Eloy find it for me. Jun, also, I'm gonna have an interface to how like, I can actually type in my own queries, Jun some kind of DSL that turns into proper SQL of what I need, as well as, of course, an MCP or command line for Skills. Speaker 5: Now, Speaker 2: a little bit of data that I've actually collected last night on some things that I found. I actually had to analyze my Claude code June conversations. So some of the interesting things AWS, resumes that are Greg with the knowledge Base. Whether it's the same session that you've compacted and you're coming back to the resume, whether it's a new session that you're coming back into the Speaker 1: project, Speaker 2: you come back in and you say, hey. I'm starting to work on this or we were here. Let's start working on it again. And, it's saying, 6.6 times more. It just reads that and finds exactly where HINTS status is Esther than if, hypothetically, I I came back into it Jun then I had to point it to all the different tracking files and all that. Speaker 2: 1 thing also that doesn't necessarily surface on this is handoff files. I very rarely have a handoff file. Almost always, I just say, update the tracker where I'm at. It updates, Check some boxes. Maybe if there's some details that are important in the session that we haven't materialized already, it'll do that. Speaker 2: And then I'm done and I walk away. And then when I come back, it can pick up right back because it's so we've kept track as we go along exactly what has been accomplished. The stored research, is definitely use. AWS as I'm sure you found, you know, storing shares the deep dives, it's easy to consult that rather than regenerating that every single time. And so there's some other information I actually Vibe, both this project dir and also the code that generated this, up on, a GitHub track actually added Antoine link on the, the thing AI I can give it to you After. Speaker 2: And I can certainly talk more about this. But, yeah. Any questions? Speaker 1: Yeah. Yeah. What's your process of iterating on this? Like, if you're as you're developing this framework, is it like, I don't know if something goes wrong and then how because it's kind of 1 Speaker 2: So you're you're talking about this bigger index thing? Speaker 1: Yeah. Or just like the structure of the file and so on. Speaker 2: So the N/A the the knowledge Base, Project structure, probably within a month and a half, 2 months easy mostly set. Like, there's been a little bit over time, but mostly I don't need more because it's it does what it I needed to really well. The whole index thing, that's gonna be a much bigger AI gonna have some failures and I just have good backups. Sean. 1 last thing is also on this directory. Speaker 2: I actually have a desktop, my laptop, and then my home server. I use, sync thing, which is open source software CTO sync it all around. So I have my desktop and my server always on, and Los server backs it up. So I then just open my laptop, connected Esther Internet. Within seconds, it's synced everything back and forth. Speaker 2: Yeah. Speaker 1: It's I Speaker 2: don't have any I don't have any controls whatsoever because it's living within my current how I actually work. Maybe there's a few things that I could do to reduce City, perhaps with some specific command line Skills, but then it's using a Base tool local, which also cost tokens. I suppose Product if it's loading up full reference documents, perhaps that that actually could be an issue. But what it tends to actually do AWS, at the time of Hooking, okay, the the on-demand the project Jun each each document file, in the index has here's AI what's in the file. It starts to Greg through and tries and finds the subsections that it needs. Speaker 0: Yeah. That was funny. Feels like it could Speaker 1: Jun context of files Speaker 0: too and then be, like, spit on the sub agent to Speaker 2: that's window of the strip that's kind of what the scratch directory can be also. Yeah. Speaker 5: Do you keep this in version control? Nope. Speaker 2: But I do have everything on ZFS with Sanoid, so I have 15 limits snapshots. Can you Speaker 1: explain that part? Speaker 2: Are you familiar with the ZFS file system? Speaker 1: Yeah. Speaker 2: So Sanoid is, basically Dataset of Chrome scripts and and Wrote scripts, s a n o I d, that you can set schedules to automatically take snapshots and prune snapshots. And you can configure per pool and per directory and all that kind of stuff. Speaker 4: AI you been, Live, I Folder like to agent that HINTS, like, an actual automated 2026? Have you been using, like, helps and, like, repetitive Esther that, like, kinda enforce the fact that you use your knowledge Base in context no matter what? Speaker 2: Not really. MoE, it's a Skills prompt in my system prompt of AI using knowledge Base. Here's where it lives. Here's the general structure. There is a Project mapping of, like, here's a name to, like, what it is and where in the Esther the knowledge base and the source code, but that's it. Speaker 2: Like, there there's and that's just a few lines for 150 line per. So So it's a a pretty basic mapping. Speaker 4: I think you should, check out, GSD. Yes. Get SSH done. And it's basically using some of those principles Speaker 1: Mhmm. Speaker 4: But Building, like, a framework around City. So you never go out of your knowledge base. Basically, you're you're kinda, like, building a project within the context of the knowledge base no matter what Jun keeping the structure of it. And so I think they're, like, interesting concept that you feel. Speaker 2: Potentially. I mean, I I haven't personally found any issues, but that doesn't mean there haven't been. I just haven't noticed them. Speaker 1: Yeah. What's the name of the GitHub? Speaker 2: Oh, AWS, agent /drizzt, capital lead. So dris 3 2 1, slash a I dash tools. And Live started putting a bunch of very random tools that I'm doing for myself that I wanna share. Speaker 1: Do you want to have dris? Speaker 2: D r I z z t, number is 321 and it's a capital d. Speaker 1: So Speaker 2: any D and D fans, you'd probably recognize the reference. Any other questions if there's time? Speaker 1: Well, I probably could relate with it so far. How how how will the current work for you for framework? Speaker 2: AI mean, it's worked pretty well. I I think of it more as a lead harness and more knowledge based framework. But AI Game gonna evolve it into the index System, which also has MQTT Event driven to ingest, have Check extract that can live locally that's a smaller subset that's accurate, but then can feedback. So when my laptop is offline, if I'm in a in a low, signal area talk can then feedback into the main that lives on my home server, it's gonna be pretty elaborate on-demand this bigger part of a bigger harness of event driven agent, automation for myself, my own personal life. Speaker 1: Thank you. Cool. Thank you. Speaker 0: Cool. I find myself setting up a similar framework every single time, so I'm just gonna take Speaker 1: Yeah. Yeah. Yeah. Speaker 2: At least. This is why I put it up on the GitHub. Cool. Speaker 7: Hi, everybody. Thanks thanks for having me. Thanks for hosting, Greg. That was an awesome hockey use. It's, like, an inspiring place to start out. Speaker 7: I'm gonna talk a little bit about synthetic time travel, which is a favorite subject of mine and how we use AIs to do that. So build a company called TimePoint, and, we focus on rendering the past to understand it and rendering the future to choose it. And I'll show MoE. Speaker 1: Oh, well, then that Speaker 7: wasn't as punchy Speaker 1: as it would Speaker 7: Live been when we did a slide up. Okay. Speaker 1: So even if Speaker 7: that takes a 2nd to connect, I'll continue talking and saying, AIs have a remarkable capacity to role play. A lot of us have gotten them to do that. Speaker 1: That's the remote. Speaker 7: I'm trying to plug this on. Alright. So AWS has HUD a remarkable capacity to role Replay, and I'll walk through a couple of Speaker 1: Go back in time. Yeah. Before AV. Speaker 7: Yeah. Before AV. Yeah. A AI would be a killer company. Speaker 1: You're like, your shit just works. Speaker 0: Yeah. Speaker 1: Our table's right. Yeah. Okay. Well Speaker 0: Oh, here we go. Speaker 7: Do you have the remote? Speaker 1: I'm Speaker 7: just Alright. Okay. Are we gonna stabilize? Okay. So I'm gonna talk about synthetic AI travel. Speaker 7: The point I was trying to make is AI agents have a remarkable ability to role play as historical figures or in historical scenes. I'm gonna talk a little bit about this new thing that we've put out called that I Vibe out called the Check chain. But 1st, I actually wanna say that, like, we're, we have a little tradition Jun Tinkers of ending with someone trying to go a little bit into the weeds on the technical side of City. And inspired by someone I met at SPEAKER last month, Leo, who's here, who Speaker 1: just graduated. Woo hoo. Speaker 7: AI he asked me this question about this graph if if it would be NP complex. And I looked at him like a deer in AI. You know, because it was AI the question AWS basically if you take network x and have it render a graph and then you give the graph to Track Jun have Grock describe the graph and give that to Claude Jun have Claude build a new graph from the SSH. What is that? And I will say I straight when, it's been on my mind so much that when Fable came out AI based. Speaker 7: And it said that essentially the question wasn't that interesting to Fable because it was set theory. Which is demeaning SSH shit if you're a human, who finds it interesting. But it easy basically saying how do you slice City? But Leo inspired me with this question. And so what I workflow all the time is, graphs and different kinds of graphs related to AI. Speaker 7: I'm not gonna walk through this, but that idea window sticks. I accurate Skills some of this. I'm gonna talk about, social network, augmented generation or snag, which is an easy way to do it. And we'll talk about the clock chain. So I'll switch out of this slide deck HUD Claude obviously made and show you the AI point web app. Speaker 7: Just 1 brief Simple. You can go into time point and say, show me the Tesla Q3 earnings call. And in this example, it AWS in a shallow mode called time point flash. So it's not going to try to calculate nickel prices Jun use know, that kind of stuff. The version time point pro does do that is capable of that kind of stuff. Speaker 7: But even if this basic version, it renders all the characters, it builds agents for all the characters. The reason I like this simple example of the q 3 earnings call is the AI window role Deploying effectively notices that the AV SSH, Mark, is a part of the earnings call. Now look, if you go ask a generic vanilla AI to simulate an earnings call, it may not mention Mark. What's important though about earnings calls is there's a linear correlation between AV quality and the stock price. Right? Speaker 7: Like, go find a Jun 50 CEO and see how seriously they take your EV AI. Very, very seriously. Replay agent backup, multiple satellite limits, 6 microphones. Elon can get on the phone and, like, call from where, you know, Bahamas, but most CEOs have so that's an interesting example where AI point teases out. It's not just investor relations. Speaker 7: It's also the a b guy in the room. Then you can use this for all sorts of stuff. You can simulate pitches. You can simulate road API. You can simulate strategies. Speaker 7: Go to market. You know, you do all sorts of stuff. What I'm gonna talk about tonight is the clock chain and the MCP access. If you go to timepointai.com/dev, you can see the MCP Wrote. And what the clock chain is is a persistent temporal graph in, like, the most dead simple terms. Speaker 7: So it uses Postgres. You can access the MCP protocol, with no auth. The idea is that you can get access to, all of Esther. And I'll just open up a web app example of, like, moment by moment. And, of course, some of the AI images are still a little, you know, wonky. Speaker 7: But what AWS slowly starting to build up is is something actually really powerful. This, you know, version of it's very SSH, but if I switch over here and go to my handy little, MTP demo tool, we can see that the clock chain, which is starting to grow autonomously, is up to about 20,000 moments with 500,000, 600,000 causal edges. And what does that mean for those of you who are not graph theory folks, a causal edge, goes back to a computer AI, Judea Pearl, who talked about causality Jun formal structures back with Danny Hillis, back in the day. Now if anybody can tell me, you know, about Danny Hillis's role in LA's, talk system, I'd love to talk MoE. Legendary LA tech guy. Speaker 7: Going back to use causal graph just means AI did this thing OpenAI? And so the clock chain is my attempt to build an open source causal graph for AI agents to be able to reliably hit. And, what that means is you could go in and just search the graph. We'll just search we've got pre based examples. Research clock chain for Apollo 11, it's still pretty agent, but we'll see, it you know, the the various AI moments that it already has with images and dialogue AWS characters Jun, critically, these causal HINTS, you know, like, the STS, 1 1 4 Discovery return flight as it relates to the Saturn 5 launch. Speaker 7: If you start to visualize how this graph starts to populate over time, it gets very big, very quickly. It also becomes incredibly powerful. I'll give you an example AWS lead say, you're trying to understand why an m and a transaction was done 5 years ago by 1 of your competitors. As the clock chain starts to grow and grow and grow, when you run that simulation to say, why did they do this m and a, Eloy, you'll be able to fold into your agent talk of the clock chain data. And what's cool about the clock chain is there's agents iterating over it 24 7 AI to further ground it, and it's starting to get pretty interesting. Speaker 7: I'll open up another moment use. We can look at the graph stats real Speaker 1: quick Speaker 7: or, hopefully, Speaker 1: lead Speaker 7: doesn't have an image for that 1. Whoops. This, s this little script that's doing these demos is something I had built, like, 10 minutes ago while someone else talk talking. But we Code see we can load up all these moments. So, the the reason I wanna talk about the Check chain, AI, and File I'll, I've got a little bit of time left, is it actually uses Esther standardized data format, which is also up on our GitHub called time point data format. Speaker 7: And what starts to happen that's really cool is when you start to give different Jun, different models, different image generators, you start to give them all the same protocols Jun the same data limits, and you start presenting the causal graph, they start kinda arguing with each Esther, and you get better and better truth. We also start to normalize the data structure in a way that's really convenient for future training data. So, I've only toyed with it, but OpenAI has a powerful tool where you can upload a bunch of data and train a small agents. So you can like train a Jun 7,000,000,000 parameter model or something and fine tune it against your Dataset oxen, because oxen plays nice from the command line time point. Use can just say, push SSH, character to AI, push my George Washington training data from the Check chain to OXN and very quickly at low Code, very efficiently agent a synthetic George Washington. Speaker 7: That's pretty high. And so my invitation to all of you is to start contributing to the clock chain. It's a, it's kind of a new idea. I do actually want to build it out like a blockchain eventually where the more your agents contribute, the more maybe payment you get for every API call that goes into it or MoE sort of distributed, collective teaching. Because the reality is most of the big tech companies already have a very high resolution knowledge graph that's wildly expensive to use. Speaker 7: Google knowledge track, there's a AI, there's lots of paid knowledge graphs, but not only are they incredibly repetitive to use, they're really not tooled for an agentic era. They don't have fine tuning parameters for dialogue and role play. So so like that some of this, you get into weird stuff. Right? With the clock chain, 1 of the things you have to do is guard against anachronisms. Speaker 7: Right? You wanna make sure that George Washington doesn't have a cell phone. Even advanced models today are still pretty bad about that. Right? Like, if you go ask Track to make a 100 pictures of George Washington, you'll see some cell phones in there. Speaker 7: It's City, you know, it it that's a very difficult thing for the models to discern. So, if you wanna see any of this or contribute, you can just go on to, the AI point GitHub and see the Flash format, which will let you render these moments Jun the, Check chain Code, and and then you can just go on the m c m p MCP server and participate in it. I think it lists all the tools. Yeah. So you can propose a moment, challenge a moment, query a moment, get a moment, or get the graph stats off from the MCP. Speaker 1: Thanks, everybody. Speaker 7: I never AI quick questions? Questions? Speaker 1: 2 questions? Yep. About, like, for businesses AWS far as to have talk window look at a business and if you wanna kinda identify City, can it, look at that kind of an industry AI? Speaker 7: Yeah. Yeah. So with the actual, like, paid Check chain app, you AI set up your business and your colleagues and your points of contact, and you you can review all that. I actually wanted to show something else, and I'm pretty sure it's not gonna work, but I'll try it. In the SSH, in the tradition of tinkers we'll try something that might workflow. Speaker 7: Greg, can I pick on OpenAI? Sure. I'm just gonna take some of the content from your website AI say Fitch Oxen. Let's see if this works. Nope. Speaker 1: It failed. Speaker 7: Okay. So what Speaker 1: I was thinking thank you. Yeah. At least it was a quick and painful mess. MoE what Speaker 7: this was supposed to do that I wanted to show everybody tonight and why I, but it's not AI a core function. It's not what I'm, like, selling to people. The idea though is I build easy synthetic venture funds and a synthetic angel investment Jun, and you CDN pitch your Esther up to City, and it will provide synthetic feedback and training for you. To your to your point, like, what yeah. The way I actually help businesses I'll give you an example of a client I worked with recently where they, they saw 1 of their competitors get acquired for, like, 10 x the the value that they understood. Speaker 7: They're like, what the you know, you're Wrote whatever. You're $10,000,000 in Avenue, and you just got acquired for $200,000,000. Speaker 2: We Speaker 7: don't get it. So what we HUD AI point do was build, synthetic dossiers of all the market lead, of all the different companies and Esther, and then we ran we just brute force it. I ran, like, 25 different simulations, something like 16,000 total API calls. And in 1 of those simulations, all the numbers spiked on what the valuation would be. I Date to be careful about what I disclosed, but we found a transaction 3 years earlier on an upstream private equity firm validating that they were doing this strategy. Speaker 7: So that may sound AI like esoteric or whatever, but it's really a fascinating competitive advantage to say, what is the competitor gonna do? Or critically, why did they do what they've already done AWS actually a very interesting thing to Sean out. Speaker 1: Yep. Are you using, public Date or working with companies like PitchBook, that kind of thing? I Speaker 7: would say API agnostic. I have not put pitch book in. I really like user, you know, paid data when it's well structured. So I teaching, yeah, if we're gonna if we're gonna double down on this, like, synthetic investment funds, absolutely pitch book. But, the time point data format, if you go to the GitHub and see it, what it actually is kind of an Esther. Speaker 7: So it would even take the pitch book data and put it in this AI of novel temporal format so that we can do causal tracking, which for what is weird 20 years ago AWS what I was doing for Shares 1 Base, if anybody remembers them back in the Venue ecosystem as building similar models. AI SSH. Shares running we want a network. Can you guys both be quick? Speaker 1: Super Yeah. Just how do you govern historical truths? Speaker 7: Use don't. We we are all participating in that, and this is my role in it. Speaker 0: Alright. If you wanna talk to Sean now. Speaker 1: Yeah. Thanks, everybody. I mean, just tell us if you need us to hang out, Speaker 0: SSH the egg rolls if we haven't yet. Put some salad over there Wrote. Unless this meat tells you where you're going. Thanks.