I got up a little earlier than I wanted to because I had a lunch meeting, but other than that it's been good. I'm very pleased. This device is to the point where I can use it as a meeting recorder now, so we need to go build up some back end to be able to get value out of the video feeds. As right now, the only thing you can do with the videos is watch them, which is not a particularly exciting interaction. I mean, you are generating some transcripts and... But it's just the audio, right? The only thing the video is used for is to generate a cool cover photo. But we can... I mean, I can go extract the highest aesthetic images per frame. I mean, to be fair, the situations I'm using it in right now, the video feed is not actually useful. Like, I've been just wearing it in conversations. All you get are people eating. But I would like to go try it at a... We're at the point where next hackathon we'll be able to use this at the hackathon and probably wind up with a bunch of videos which are out of focus and at the wrong angle. But at least we'll learn how to tweak the focus algorithm and the angle of the camera. I mean, the angle's not a big problem. I can just put a piece of foam on it to hang it like this and it'll be fine. Yeah, I mean, I would like to ask what drinks are people drinking or on the cams or... All these features and more could be yours if you implement them. You're an engineer. I'll build up... I don't have a second one today but I'll go build up a second copy with the good firmware for you. Oh, okay. That would be amazing. Yeah. Those have bad batteries, so they don't work. But the white ones have working batteries. Oh, nice. Yeah, that was a good little assessment pitch. Oh, this could be yours if you engineer it. Yeah, yeah. Like, as an engineer, you too could have all the features you wanted. I mean, I was waiting for this moment. I mean, I cannot wait for this. It's really nice to have the device because once you start using it, you immediately start... Like, you start by implementing the features that you think would be useful. Yes. And by doing... By tinkering with this for a while, you get a much better idea of what the back-end needs. Yes. Absolutely. And then secondly, did Brenda come by? She didn't. She sent me a text saying she was busy painting Halloween decorations and couldn't make it. Yeah. I already spoke to her. Like, yeah, just contact them and... Yeah. I mean, it's fine. I should have... Let me go message my other guy. I should have another guy coming in today to build furniture. I was going to tell him... I told him, like, you can probably help them organize, like, a garage sale or something. They got a lot of access to furniture. Yeah. I have another guy coming here to build furniture. I've been trying to find a fresh, like, persistent task rabbit. The... My original guy who, like, worked with us for a couple of years, I told you, right, he found a six-figure job as the garbage man for the commercial districts of San Francisco. Apparently, the night garbage man starts at 100k... Starts at 100k a year and goes up to, like, 150, which is a pretty solid deal. I mean, it requires a CDL, so it is a high-value job, but it's less stressful than being a truck driver. Then my second task rabbit, my current theory is that he either... Hopefully, he just dropped his phone, but he may also have dropped himself. It's unclear. One ominous Saturday, he told me I could... Can't make it today. I'm working a construction site that I never heard from him again. So I'm a little worried. He was a good guy. The problem with having a large house is, like, like, we have so much acreage and all we do is work and meet in the entryway, like, literally the entryway. You see, normal houses don't have entryways. They go straight into, like, a family room, but no. Good mansions have entryways. But because of the lack of entryways, it's hard to find a job. I mean, it's hard to find a job. But no, good mansions have entryways. But because they're close to the front door, you just kind of sit down and work in them. Also, that freaking El Granada house. Holy crap. That thing's incredible. Yeah, the one with the gigantic chandelier. Yeah, it has a 15-foot pole. It has, like, a two-story chandelier. Someone got a steal for that house. Holy 10 mil. Like, hot dang. Hot dang. Yeah, holy crap. You have thrown a lot of parties, so they know how to turn this into a party house. I mean, I live in an exotic right now with neighbors that live in similarly exotic houses, so I know that these large houses are impractical, but every time I see one, I think about moving into it. I like, like, how that thing had 11 bedrooms, and each, like, level was, like, different carpet, different vibe. Like, three separate pools, all of them not designed for, like, swimming. It's all designed for being close to each other. Yeah, it's also 23,000 square feet. That's a big house. Yeah, it's a lot. I guess the size of AGI house. It's like you're in a B-2 bomber. Yeah. All the way up. And it's elevated on 11 rooms, and all the rooms are just different, and it's got this, like, 70s, 80s, sorry, 80s aesthetic to it. All right. I guess, I don't know what we want to talk about first. I guess, from my point of view, there are two clear things that I want to do. One is get value out of the video stream. I don't know what value that looks like. The original plan was to go convert the video stream into images and rag it. That's what is ready. What I want to do is convert the video stream into images, then somehow pair those images against those transcriptions to get a little more context. Can you get a USB-C charger? Yeah. Does this reach? Yeah, plug it into that tower. So I think that would be useful. Because, like, take the videos we saw today or, like, the videos at this meeting. I also want to really do better diarization using the video feed, which I think is possible. I mean, is this a, are we sort of indexing? You're already at the stage where you just, the videos uploaded are somehow in the backend. And then now we're in the process of figuring out how to index the process. So Raj has a demo where you could upload a video. It does, I think, it does inter-frame similarity in order to detect frames with large changes. Then picks the waypoints which have the largest changes. And uses, then turns each of those into image records and saves them for future retrieval, your flashlights. Chapters in a book. Yeah, chapters in a book. Then right now under the, well, I haven't deployed in the backend yet. But right now in the current implementation, all of those images get stored in the same database that, like, user photos get, like, user-triggered photos get stored in. But I think I want to change that. I think I, in Merit's discussion, actually, now that we're all here, the other way to do it would be to build, like, a little mini VectorDB for You're not wrong with that. If the database that you try to do supports very good filtering. Yeah. And categorization like this comes from this. You're completely right on that. It may not be the right design. The other design would be to go cluster data from what we believe are semantically related recordings into one database to, like, help the search. These are sort of the same thing. Yeah, but the one thing you want to avoid is, in the future, if all you have is effectively, like, video data, and your question comes in, and you're like, well, our video data RAG system, I guess...