Doomsday Stick

Speaker 0: I'm here to talk to you about the end times, doomsday, the apocalypse, and I'm here with a solution. In your pocket, you can carry all the knowledge of mankind, all the knowledge that can fit in 31,000,000,000 parameters. I am here to, share with you a project, mostly AI coded. It is called the Doomsday Stick. The Doomsday Stick is a Gemma4 for 3100000000 Fair bit quantized model crammed onto a 32 it's only 17 GIVE, so crammed onto a 32 gig thumb drive. Speaker 0: And it is meant to run offline, on app specific Apple introduces. They're capable of running the MLX meeting the specs and running Fair MLX platform local local Science. 2026 its goal, if I plug this in if I plug it in, AI GIVE the doomsday drive, but that's just my thumb AI, and I have 1 single DMG. So this is for running on MacOS. So it's a full on Swift application. Speaker 0: I have never written in Swift before, and I Stick, to this day, Time proud to say I have not. I'm a Flutter dev. That's how I live. But this is a direct native Swift app, that I built. The flow that I used on it was, hey. Speaker 0: I lead this I had this idea. I released a video about Apple did WWDC, and they featured the WLX WLX platform for running local, and then also their core AI Fair supposed to be used by my normal applications to run local inference. I got halfway through. After 2 hours, it just was crashing on me. But I did get MLX working, enough to see some of the promise on it. Speaker 0: So that's where it came from. It's like, Date. I could with this capability, I could make a more normal type of app. I could develop it proper style, and then only use local inference. I got the SIGN. Speaker 0: I'm like, this is a brilliant idea. I got the design for it. I went 2026 AI went into Claude chat, and I said, hey. Give me a design for a, a post apocalyptic Stick CRT style cartoony electronic interface, was my inspiration. I told it to give me a AI, and I expected it to do a a design file that I could work off of. Speaker 0: It did a a a single, HTML standalone file that was actually animated. Like, it had everything in it. It was it was gorgeous. Took that, threw it into Cursor. I had Opus 4.8 write up the plan Found then execute it, make a shell of the Cropped, a shell of the app, which looked perfect. Speaker 0: And then it popped up on Cursor, and it was like, hey. Fable 5 is out. You wanna use that? I'm like, I don't even know what that is yet. Yes. Speaker 0: So I had it file up the plan to use the Local Jun and hook it up and set a basic chat, chat functionality, and it 1 shot at it. But it chewed through tokens on it, but it came out beautiful. So everything is bundled up into a Stage alone DMG for Apple. And within it, if I open it again off the drive, we have the application itself and then the model, which, as I said, was a Gemma4 model. So this is all the weights, compiled down that City next to it. Speaker 0: The 2 of them together is about 17 GIVE, and you can run it right off of the Drink, but it's slow to boot up. I would recommend dragging it to your desktop, which is also slow. And that's why we're gonna AI LLMs the exact same file over here Found run it from here, which fires up rather quickly. I should add resizability, but you get the cute aesthetic on it. We see here that it's acquiring model, Gemma4, blah local blah. Speaker 0: It's doing its fancy UI thing because it's firing up in the Found, the MLX instance. And here we see Gemma4-31B-Q4 Found the memory Stage. Is running just under 18 gigabytes of RAM. I should GIVE pointed out, I have an m 4 Mac with 48 gig of RAM. You should only need 24 gig Jun any m series processor, and you too can have a survival instrument and knowledge base at your fingertips. Speaker 0: So say, oh, no. It's an EMP. We've lost talk signal and data. This is still Building. 100% local. Speaker 0: I have some examples AI, how do I convert a car to run on, what is City, wood gas? Well, the model knows about it, and it's cranking out AI here 10 Building 11 tokens a 2nd on my my local off-line. It's a it's a denser 1, which means it has a fair bit of world knowledge baked into it. So this is legitimately telling me how to build and run a gasification conversion to my car if Tier where you have no petroleum lead. Food safety, 1st aid, and, of course, all the other knowledge of the world. Speaker 0: If you wanna recreate the works of Shakespeare, it probably has most of them memorized. He's not he's probably gonna sue Time for that. So what do you got? You're in you're in the apocalypse. What information do you need that, to look up that you don't know about? Speaker 0: I have Speaker 1: purified water. Speaker 0: That's 1 that I Speaker 1: did before. Speaker 0: How, what I what I was thinking was how to purify water without heat. And it's all really good stuff. Best practices, multiple methods, and, of course, all of this is coming from that 17 gigabytes of weights, which is crazy. Next step, spoiler alert, it turns out it's not that many gig to get, like, a substantial chunk of the text from Wikipedia and throw that on there Food. So I might be adding that to try and enhance it. Speaker 0: But there is it's an insane reality that we live in that so much knowledge can be crammed into 17 gigabytes of of data because it knows all the things, not just survival, but all of it. There is a lightweight prompt on top of this. How should I go about finding a bunker? I don't know. Let's find out. Speaker 0: Finding and evaluating safe shelter, site selection, high Found, always do the high ground, identifying existing structures, Food cellars, Folder war era bunkers, structural integrity, ventilation, so it gives you safety. It's just so good. I've been so impressed. Like, this was on a whim, and I'm like, it it should be able to do some HINTS, but the way it breaks it down and gives you such a gamut of things, different options, and cautionary things. But it's not guardrailed. Speaker 0: You know? It's not a pain at the end of City, but it does say, like, hey. Take these potential measures of safety with let me show you this 1. What what animal did you just catch? Speaker 1: Yeah. Speaker 0: Squirrel. Oh, how many calories are in a squirrel? And I don't know if it's if it's imagining it, but it breaks it down. I did this for a rabbit, and it breaks it down. Speaker 1: For me too. Speaker 2: Like, what Speaker 1: what plants should I collect? Speaker 0: That's part of that that's part of the, sales pitch on the website. You know, I went for the lead slash prepper population and the overlap thereof, but part of that City kept coming up with was, plant identification, with a big disclaimer, like, don't just use this because you'll probably die. But if it's it's better than nothing, that's kind of the the thing it has. It could definitely tell you how to use I'm sure you know, how how do I June a compass to navigate? Speaker 1: Date you don't know where you are or if you know about the towns nearby, do you find out details about what's in that town, like the stuff that's on Google Maps? Speaker 0: Oh, yeah. Yeah. That's very large. That'd be but you're you're much more, hey. How do I I have a compass and a map, and I have no idea how to do it. Speaker 0: But this is a full full chat. So if there's any, of course, any point that you don't you know, what the hell are you talking about on declination? Speaker 1: Just Speaker 0: ask Speaker 1: things like just to check if it's local knowledge, ask if which Speaker 0: City, let's see. Let's see. What cities are close to Binghamton, New York? I'm curious. Endicott, Johnson City, Vestal, Owego, Scranton, approximately Tier, 60 ladies and gentlemen, did I point out that this came off of a thumb drive Found that's all baked in there? Speaker 1: Did you ask it how to get to 1 of those places from Speaker 0: let's find out. How, yeah. How do I get from Bingham to where? Where are we going? Scranton? Speaker 0: Primary route Fair take Tier 81 South. Continue I 81 South approximately 50, 60 miles. Exit towards Scranton. Alternate route, New York 17. Ladies, let's clap for Time the little gemma here. Speaker 0: And this is why I chose a you know, it was recommended to choose a dense model, the the dent instead of a mixture of experts, which is more efficient. But for things like this where you want it to have baked in and compressed all of human knowledge, a dense model is where it's at. And I've just I've just been blown away with this. The the requirement to run it is impressive that it's here. You know, granted, this was a $2,500 machine last year. Speaker 0: But the fact that we're already here on it Found they tend to get more efficient, and with the right iPhone, you can run the the same inference on it. It's just gonna be a smaller model. As you saw, this is a legitimate product for sale that you can get at doomsdaystick.com. It's a great landing page that positions it for everyone along with the spec sheet Found questions AI, do I get updates? No. Speaker 0: Its whole purpose is to be offline. Yeah. So AI, already the next Future lined up. Hopefully, a whole navigable intro web, offline version of most of the text or at least a good chunk of the text to Wikipedia is next for computers that aren't capable, there's still a lot of value in having that resource where you can plug it in and navigate the links directly. And then if you, you know, stumble HINTS kill enough zombies and get into Best Buy and get a better computer, then you have the ability to have the world's knowledge at your fingertips.