Introduction to On Device AI(Learn to Code Tuesdays) 🔗 **Source:** [Introduction to On Device AI(Learn to Code Tuesdays)](https://www.youtube.com/watch?v=UyspFz7DEOQ) > [!note]+ Document Notes > tags: [ai, ondevice, machine-learning] > author: Techoploois > source: YouTube > reference: https://www.youtube.com/watch?v=UyspFz7DEOQ > date created: 2025-01-15 10:30 > > # 🏷️ Title: Introduction to On Device AI (Learn to Code Tuesdays) > <!-- deep-read:sections start --> ## 🧰 Action Inbox - [H1] Install Ollama from ollama.com and test with llama3.2 model — 00:16:55 #setup - [H2] Check RAM capacity to determine appropriate model size — 00:26:30 #hardware - [H3] Download Perspective AI app when released for Apple Intelligence — 00:11:20 #tools ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - [Ollama](https://ollama.com) - [Techopoulos](https://techopoulos.com) ## 📝 Summary - Tutorial on on-device AI models that run locally without cloud connections - Covers model parameters, RAM requirements, and installation process - Demonstrates Ollama terminal interface and model testing - Discusses upcoming accessible tools like Perspective Studio - Shows practical vision model capabilities for image description ## 📖 Main Content On-device AI represents models that run locally on your hardware rather than requiring cloud servers. Unlike ChatGPT or Claude which query remote data centers, these models work offline after initial download. The technology addresses privacy, cost, and environmental concerns while enabling AI functionality without internet dependency. Models are measured in parameters (3B = 3 billion, 70M = 70 million) which correlate roughly to RAM requirements. Generally, models under 1B parameters need ~1GB RAM, while 8B models may require 8-10GB. Storage size doesn't directly equal RAM usage, and quantization can compress models at potential quality cost. The demonstration shows Ollama, a Meta-backed platform for downloading and running models via terminal commands. After installing Ollama, users can pull models like llama3.2 or gemma3, then interact through command-line interfaces. Vision-capable models can describe images with surprising accuracy. ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=0) — [H1] Introduction and Program Transition - Key idea: Laura Wheeler Foundation transitioning program to Techopoulos in 2026 - Evidence: "There's only one of me and you know, we're not a big organization" #### ⏱️ [00:03:45](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=225) — [H1] What is On-Device AI? - Key idea: Models run locally without cloud servers or internet connection - Evidence: "It's basically a model that can work offline. It does not need a data connection" #### ⏱️ [00:05:20](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=320) — [H2] Cloud vs On-Device Comparison - Key idea: Cloud AI uses remote data centers while on-device runs locally - Evidence: "There's a data center where the models are on a server somewhere" #### ⏱️ [00:07:15](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=435) — [H2] Model Parameters Explained - Key idea: Parameters indicate model size - B means billion, M means million - Evidence: "A parameter is basically like how big the model is and like how much data and other things that it can hold" #### ⏱️ [00:08:30](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=510) — [H2] RAM Requirements Guidelines - Key idea: Model parameters roughly correlate to RAM usage - Evidence: "If the model is like less than a billion parameters, it's probably going to take like one gig of RAM or maybe less" #### ⏱️ [00:09:45](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=585) — [H3] Multimodality Support - Key idea: Models support different input/output types beyond text - Evidence: "Some will support image while some will support video and some will support other different kinds of outputs" #### ⏱️ [00:10:30](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=630) — [H2] Context Window Limitations - Key idea: Context window determines how much data model remembers before forgetting - Evidence: "The context window refers to the amount of data a model can hold before it forgets" #### ⏱️ [00:11:20](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=680) — [H1] Perspective Studio Announcement - Key idea: Accessible alternative to LM Studio coming to app stores - Evidence: "We're going to actually open source this... it's going to be called the perspective studio" #### ⏱️ [00:13:15](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=795) — [H1] Apple Intelligence Integration - Key idea: iPhone users already have on-device AI through Apple Intelligence - Evidence: "Well, we're actually making an app... that uses the Apple Foundation models" #### ⏱️ [00:14:45](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=885) — [H1] Model Aggregators Overview - Key idea: Applications that manage multiple models from different providers - Evidence: "Open Web UI... LM Studio... those are the main applications" #### ⏱️ [00:16:55](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=1015) — [H1] Ollama Installation Demo - Key idea: Step-by-step process to install and run Ollama on Mac - Evidence: "Go to ollama.com... install for your platform" #### ⏱️ [00:21:30](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=1290) — [H2] Terminal Model Interaction - Key idea: Command-line interface for running AI models locally - Evidence: "ollama run llama3.2... when you hear this greater than greater than greater than you're good to type" #### ⏱️ [00:26:30](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=1590) — [H2] Hardware Compatibility Guidelines - Key idea: Match model size to available RAM for optimal performance - Evidence: "If you got 16 gigs of RAM, you could probably go with with an eight or maybe a 12" #### ⏱️ [00:35:20](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=2120) — [H3] Quantization Concept - Key idea: Compression technique that reduces model size but may impact quality - Evidence: "Quantization in short terms means that basically it kind of compresses the model almost like a zip file" #### ⏱️ [00:52:15](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=3135) — [H1] Vision Model Demonstration - Key idea: On-device models can accurately describe images - Evidence: "The screenshot shows the main menu of the Blizzard mobile game Hearthstone" ## 📚 References - [Ollama](https://ollama.com) - [Hugging Face](https://huggingface.co) - [LM Studio](https://lmstudio.ai) - [Open Web UI](https://openwebui.com) ## 🗣️ Paraphrase - On-device AI runs locally on your computer without needing internet, offering privacy and cost benefits over cloud-based services - Model size is measured in parameters (billions or millions) which roughly indicates RAM requirements for running the model - Installation involves downloading Ollama, pulling desired models, and interacting through terminal commands or GUI applications ## 📰 Newsletter Exciting developments in accessible AI! On-device AI models are becoming more practical for everyday developers, running locally without cloud dependencies. This technology offers privacy, cost savings, and offline functionality. Key highlights: Ollama provides easy model management, vision models can describe images accurately, and upcoming tools like Perspective Studio will make this technology more accessible to screen reader users. The demonstration showed how to install and run models locally, with practical guidance on matching model size to hardware capabilities. Perfect for developers wanting to experiment with AI without cloud costs. Resources: [Ollama](https://ollama.com), [Techopoulos](https://techopoulos.com), [Apple Intelligence](https://apple.com/intelligence) ## ✅ Action Items #### DevRel - Developer advocacy: Create accessible tutorials for on-device AI setup - Education: Develop workshops on model selection and hardware requirements - Developer experience: Document terminal accessibility for screen readers - Community management: Foster discussions about on-device AI applications #### Community - Contributors: Support open-source model aggregator development - Maintainers: Provide feedback on accessibility features in AI tools - Sponsors: Fund development of accessible AI applications #### WordPress - Core AI: Explore on-device model integration for content generation - Accessibility: Ensure AI features work with screen readers - Documentation: Create guides for local AI model deployment ## 🧪 Hypotheses - On-device AI adoption will increase as models become more efficient and hardware improves - Accessible AI tools will democratize AI development for disabled developers - Vision models will become standard for accessibility applications - Local AI will reduce cloud dependency and associated costs ### 📊 Methodology - Compare cloud vs on-device AI performance metrics - Test accessibility features across different model aggregators - Measure RAM usage patterns for various model sizes ### ❓ Questions - How will quantization affect model accuracy for accessibility use cases? - What hardware specifications will be standard for on-device AI in 2025? - Can on-device models match cloud model accuracy for specialized tasks? ### 🔮 Future Research - Benchmark accessibility features across AI model platforms - Investigate energy efficiency of on-device vs cloud AI - Explore integration possibilities with assistive technologies ### 💡 Implications - Practice: Developers can experiment with AI without cloud costs or privacy concerns - Policy: Local AI reduces data sharing and privacy compliance complexity - Security: On-device processing eliminates cloud-based data exposure risks - Education: Accessible AI tools enable broader participation in AI development - Funding: Open-source AI tools reduce barrier to entry for small organizations ### 🧠 Learning Styles Assessment Reading level estimate: 12th grade #### 👀 Visual learners - Present: Terminal screenshots, model parameter comparisons - Improve: Architecture diagrams showing on-device vs cloud workflows #### 🔊 Aural learners - Present: Live demonstration with screen reader narration - Improve: Audio-only tutorial versions for complex installation steps #### 🗣️ Verbal learners - Present: Technical term definitions, clear analogies for parameters - Improve: Glossary of AI terminology, comparison frameworks #### 🧩 Kinesthetic learners - Present: Step-by-step installation guide, hands-on terminal commands - Improve: Interactive checklists, sandbox environments #### 🧠 Logical learners - Present: Parameter-to-RAM correlation tables, decision trees for model selection - Improve: Performance benchmarks, compatibility matrices #### 🧑‍🤝‍🧑 Social learners - Present: Live Q&A session, community discussion - Improve: Peer collaboration guides, shared experimentation spaces #### 🧘‍♀️ Solitary learners - Present: Self-paced tutorial, individual exploration time - Improve: Personal progress tracking, reflection prompts #### 🌿 Naturalistic learners - Present: Environmental impact discussion of cloud vs local AI - Improve: Energy consumption comparisons, sustainability considerations ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading structure, detailed narration of terminal output - Improve: Semantic markup for code examples, structured navigation #### Keyboard navigation - Present: Terminal-based interface supports keyboard-only operation - Improve: GUI alternative planning, skip links in documentation #### Contrast and readability - Present: High contrast terminal display - Improve: Customizable color schemes, font size options #### Text resizing and reflow - Present: Terminal text scales appropriately - Improve: Responsive documentation layout #### Media and captions - Present: Live narration of screen actions - Improve: Post-production captions, transcript availability #### Forms and interactions - Present: Command-line interface with clear prompts - Improve: Error message accessibility, input validation feedback ### 🌍 DEIB Assessment #### Cultural breadth - Present: Global accessibility focus, open-source emphasis - Improve: Multi-language model examples, international hardware considerations #### Inclusive language - Present: Plain language explanations, defined technical terms - Improve: Avoid assumptions about hardware access, economic barriers #### Neurodiversity support - Present: Structured presentation, clear step-by-step process - Improve: Multiple explanation formats, processing time considerations #### Representation - Present: Accessibility-first approach to AI tooling - Improve: Diverse use case examples, varied application scenarios #### Psychological safety - Present: Encouraging tone, "it's okay to experiment" messaging - Improve: Explicit failure-friendly language, troubleshooting support ### ✅ Logical Validity #### Strong reasoning - Deductive strength: Clear parameter-to-RAM correlations established - Inductive strength: Hardware recommendations based on testing experience - Causal reasoning: Links model size to performance requirements effectively - Analogical strength: Good use of "zip file" analogy for quantization #### Fallacies - Hasty generalization: Some RAM estimates lack precision for edge cases - Appeal to novelty: Emphasis on "latest" models without quality comparison #### Tone of voice - Overall tone: Enthusiastic and encouraging, with practical focus on accessibility ### 🛡️ DARVO Analysis - Deny: No concerning denial patterns observed - Attack: No attacks on critics or alternative approaches - Reverse victim/offender: No victim/offender role reversals present - Emotional cues and rhetoric: Positive enthusiasm without manipulation ## 🎯 Conclusion On-device AI represents a significant shift toward accessible, privacy-conscious AI development. The demonstration of Ollama and vision models shows practical applications for developers, especially those using assistive technologies. The upcoming Perspective Studio promises to make this technology more accessible to screen reader users, potentially democratizing AI development across disability communities. ## 🧱 Refactor Appendix ### 📚 Literature Split Pad Purpose: Stage CEE units that become Literature notes. Split level: H3 ### [H1] On-device AI models run without cloud dependency > [!quote] [Introduction to On Device AI](https://www.youtube.com/watch?v=UyspFz7DEOQ) by [[Techoploois]] via [[YouTube]]. Published 2024. Accessed 2025-01-15. > "It's basically a model that can work offline. It does not need a data connection." Establishes fundamental distinction between local and cloud-based AI processing. [view highlight](https://www.youtube.com/watch?v=UyspFz7DEOQ&t=320#H1) #xref #articleseed ID:rw_ondevice_01 ### [H2] <!-- deep-read:sections end --> ## 📺 Video ![](https://www.youtube.com/watch?v=UyspFz7DEOQ) --- # 📄 Full Transcript *Auto-generated from YouTube captions · timestamps roughly every 30 seconds · [watch on YouTube](https://www.youtube.com/watch?v=UyspFz7DEOQ)* **[00:02]** Anyway, we're live streaming now. So, >> we are we're doing introduction to ondevice AI today. Um, we're going to talk about it. >> We're going to get into the very nuts and bolts of it. We're going to have fun. Uh, it's going to be only an hour today because I've got a important work meeting I have to go to. Can't get out of it, unfortunately. Um, but it's fine. Um, so we're going to talk about some things. Um, yeah. So, introduction to on device AI **[00:33]** is our topic today. >> Absolutely. >> So, me go promote this real quick. >> We only have about six weeks left. We do. Yeah. >> I guess it's eight or nine or seven. I don't know. >> Basically two months. **[01:04]** >> We're getting to the end of the year. So, I think eventually we're are we going to have a whole uh talk about the whole Well, we've already figured it out, but I think we're going to embrace people uh for what's going to happen next year because we are going to uh have it at the same date and time for now. >> There we go. Techoploolis, for those who are watching who don't know, Techopoulos is going to take this over. Um, because while the Laura Wheeler Foundation is extremely proud and happy of all that we've accomplished, there's only one of me and you know, we're not a big **[01:36]** organization. While we're very grateful for our supporters and friends, you know, we don't have like 50 volunteers or anything crazy. It's me and my family primarily and those who can at times come alongside. So there's only one of me which is what I'm saying and I need to focus on you know other things next year and I've given all I have to give towards this. I am super proud of all that we've **[02:07]** accomplished, but we've moved way way beyond my coding skills and my knowledge and my ability to contribute. So, I'm letting Tech Techoploolis take it over. Hey, can you get rid of all the AI bots? >> Yeah, I tried to and they keep coming back. >> I know. It's very frustrating. Um >> Um >> but anyway, guys, so Techopoulos will be taking this over in 2026. I'm really excited about all that you're going to accomplish and everything that you're **[02:38]** going to do with it. Um, and where this is headed. I think it's great. We're proud to have been a partner in it and we can't wait to see what will happen next. So, >> yes. So, >> yeah, >> real quick. I'm just trying to um just get some things real quick sorted out real quick and then we'll talk about introduction to ondevice AI. This is always something I like to talk about. So >> absolutely. **[03:09]** >> All right. TGBT is having problems, I think. So I'm trying to sort that out real quick. Hold on. Let me And I know nothing about ondevice AI except for on Apple. But I this is a topic of interest to me because I'm interested in like how other models can be run on device. Um and for what kinds of application **[03:41]** whether it be whether it be you know coding or whether it be picture or video making or whether it be >> desri you know describing things or whether it be game play or writing tools or whatever it might be. It's interesting to me you know. >> Yeah. Yeah. So, let's get started. I don't want to waste time because we have a little less time than we normally do because of work means. So, uh yeah. Hello everyone. I'm Taylor from Techopoulos and today we're doing an introduction to on device AI. We're talking about what it is, how it can be **[04:13]** used, different things to keep in mind. So, this is going to be kind of a part one of part two. So, next week hopefully, I'm praying that we have a full block of time and then we'll really get into the practical applications. But before we get into the practical stuff, I really want to take a little bit of time to understand, you know, how it works and so that way you guys can talk the language in a sense because before you actually try the ondevice AI, you need to actually know what it is capable of and how it works and why you should **[04:45]** use it and maybe why you shouldn't use it. That's what we're going to be focusing on today. So, really quickly again, if you have questions in the chat, go ahead and uh put them in YouTube. Um, and then I put the information if you want to join us on stream, you could do that as well. Um, and yeah, so that's what we're doing. Um, very excited about this. So, let me quickly get a couple things in order and then I **[05:16]** will begin with the first topic. So, let's see really fast if I can get this working here. Okay. Um, all right. So, let me quickly I'm just trying to get um There we go. Okay. So, let's start with **[05:49]** Oh, really? Okay. Sorry, this is being stupid. Okay, we are now ready. So, if I can get this stuff out of my way real quick. Um, so let's start off with a simple question, I guess, or a statement. What do you guys think I device AI is? Right? Maybe pretty obvious, but what do you think it is? Anybody know? This is **[06:20]** just a question for the chat or for the Zoom participants. What is on device AI? >> Models that run without having to go run on a server or some different platform. They're running on your device. >> They are. Yes. Now, so we understand that models are running on your device. They don't run on the cloud. So there are two types of AI model you know sorry two types of places where AI models can run. There is the cloud **[06:51]** which everybody's familiar with. Think about chat GPT cla Gemini a bunch of others. Those are running on the cloud. So how those work is there's a data center where the models are on a server somewhere and when you ask it a question it will query that data center in the model and give you a response but it's using like a lot of power and a lot of energy and other things in that data center. Okay. So there's many different data centers around the world and **[07:23]** >> hey Alfred the big labs use them actually. Well, all cloud providers use them and they can be very very energy inefficient, right? Because people are having to, you know, you know, how many times you query AI every day, right? So that's something that people are always concerned about is the environmental impact. We'll talk about that later. So yeah, cloud AI runs on the cloud. It is in a data center in some location all over the world, right? Maybe in like, you know, your country or may not be, **[07:54]** who knows? And the other thing to know is that the ondevice. So now we're talking about ondevice AI. So we know that cloud AI works on the cloud. So basically it will go to a data center somewhere, right? We understand that. However, how does ondevice AI work? Well, ondevice AI works kind of if you think about like an offline map or an offline something. **[08:25]** It's basically a model that can work offline. It does not need a data connection. It does not need anything. Now, there is one exception to this rule and let's say you're downloading a model. So, ondevice AI can either come in two forms. They can come in where you download the model. Think about things like we we'll talk about this more next week, LM Studio or Open Web UI or others where you actually download the models or they can come in another form where **[08:55]** they're bundled models. The best example of this is the Apple models that are released with Apple Intelligence, the Apple Foundation models. So, the only time if you're using the downloaded models, you'll need an internet connection is when you're actually downloading the models. Who are these models good for? Well, they're good for anybody who wants to save costs because remember they're free. They're also good for anybody who's curious and concerned about the environment. They're also anybody who wants to tinker and see **[09:26]** what they can do on their device. And they're also good for maybe if you're in a situation where the internet may not be fully available or it's slow, you could still interact with AI. Okay, so those are the primary situations where people would be interested or you're just a tinkerer and you want to see what's possible. Okay, so that's also another reason. Now, another thing to keep in mind is that these may not be as accurate as the cloud models. So these models come in what are called parameters. Okay, so a parameter is **[10:00]** basically like how big the model is and like how much data and other things that it can hold. Okay. So for example, if we take the Apple uh foundation models, it's a 3B parameter. So if you've ever heard this B, this you know parameters before, what that means is it's three billion. Okay, cuz it's, you know, 3 billion parameters. Now some of them might be like 70 billion or they might abbreviate it to say 70B which **[10:32]** basically means 70 billion parameters. The other thing is is that some of them could be like 70 m. You're thinking 70 m? What is that? It's actually million parameters. So that's even smaller than like a billion parameters. Okay. And then you're thinking, well, how much resources does this take? That's a good question. So the resource limits depend on how big the model is. So generally, again, this all depends on the model itself, but here are some good guidelines to follow. The first thing is **[11:05]** that if the model is like less than a billion parameters, right, it's probably going to take like one gig of RAM or maybe less. Okay, so that that's really interesting to note. If it's 1B or 1 billion parameters, it'll probably take either one to two gigs. And then if it's uh two B parameters, it could take like four gigs and it can go up and up from there. So the other thing to note is that if you download the model, the storage, so like it'll say like this model is 3 GB does **[11:37]** not always correspond to the RAM. Okay, the RAM on your computer that means. Okay, so that's something to note as well. With that all being said, there's also other things to keep in mind if you're trying to find a good model for your device, and we'll get into that in just a second. But really what you have to know is that models come in all shapes and sizes. They come in different forms. So there's something in the models, it's called multimodality. **[12:07]** And what that means is what the model will support. Most models if not all of them support text. Okay, that's just your basic run-of-the-mill output input and output. Okay, text. Some will support image while some will support video and some will support other different kinds of outputs. Okay, so what you know you can create images or create other things with some of these models. Okay, but the most popular one by far is text. Okay, so you're probably thinking this all sounds great. **[12:39]** All right, this really, you know, it's great. It's awesome. It's what I want to do. But you're thinking what do I need to do though? like how big can these models be in the context window. So there's another thing you have to keep in mind and that's called the context window. That's basically uh between the input and output of like you know so let's see 4,000 characters. So basically it's like 4,000 uh tokens. So a token is simply it could be like two or three characters or almost a whole word. It **[13:09]** just depends. So what you're thinking about here is how big is this context window? So the context window refers to the amount of data a model can hold before it forgets. Okay, amount of data the model can hold before it forgets. Something that's very important because if you go over that contact window, you may have seen with GPT, if you ask it something that you told it earlier, you'll notice it forgets. Okay? Have you guys had that before? You ask GP other **[13:39]** models. Yeah. >> Yeah. That's because the context window has been exhausted. Okay. And some models will even say like summarizing. And if it says summarizing, you know that it's pretty much almost done. So you may want to start a new chat. Okay. But this is more of an issue on ondevice AI models because some of the context window is not as big as what you'd see in the cloud models. Okay. So that's kind of something to keep in mind. Okay. So, um, **[14:12]** what do you guys do you guys have any questions so far? What do you guys think? >> I think this is really cool. And it's I always I was trying to figure out the different models and stuff on my own. And so, it's good to talk about like the millions or billions of parameters, how they note that, and like what all this stuff really means, you know? >> Yeah. And uh something else um Techoploolis is coming out with something that's really cool and I'm super excited. We're going to actually open source this. Um but we're also **[14:44]** going to make it in the app store. So if people want to purchase it, it's going to be like two or three bucks. Um it's going to be called the perspective studio. And what this is is like the LM Studio I've talked about earlier where you can run models on your device and see if they will work with your amount of RAM and other things. But the problem with LM Studio is that's not accessible. The perspective studio will be fully accessible and it will eventually be coming to Windows as well. So that's something to note if you guys **[15:14]** are interested in that. Again, it's going to be really cheap and you can also, if you don't want to pay for it, you can download the code on GitHub and put it in Xcode and run it for yourself. But we want to at least make some profit off of this because it's it's going to be open source. People can learn from the code, but if they want the full finished product, they can or not finished product. If they want the full like app, they can go purchase it for like two or three bucks. Pretty cheap, one time, no subscription. So, yeah, that's going to be coming out **[15:45]** in the next few weeks. I'm very excited about it and you guys um will definitely like it because it will say for example with one of these models, hey this one's not compatible with your device or this one works well on your device or this one might be slow on your device. So it takes the guesswork out for you. Okay, so that's something to keep in mind. Now, let's talk about other things **[16:15]** as well. So, did you turn off the waiting room, Laurel? >> I did. >> Yeah, we've had Alfredo and David join and Yeah. >> All right. Well, I have no alerts on, so thank you for letting me know that. >> Yeah. >> So, um, do you guys have any questions before I continue? Let's see. >> I'd say so far so good. >> Yes, I think so. **[16:52]** So, let's see. Now, we're going to talk about something else with ondevice AI. And this is what you've been waiting for. How do I choose the right model or what models are good? What model should I stay away from? That's a good question. So, there's a few different models that you can choose. Actually, well, there's actually a lot. And then I'm also going to cover **[17:24]** how you get models, okay? Cuz you're like, well, great. This all sounds fantastic. How do I get them? So, we'll start with the most basic level. The most basic level is that let's say you're using your iPhone and you have the newest Apple Intelligence features. Okay. Well, we're actually making an app and it's almost ready for release. We're excited. It's called Perspective AI and that uses the Apple Foundation models, the new models that they came out with this year. And congratulations, you're **[17:55]** already using ondevice AI. Isn't that awesome? And it automatically works with your system. There's no configuration needed. It's just like plugandplay. Boom, boom, boom. All right. So, that's where I'd first start. Start with the built-in models in your platform. In the case of like Windows or other things, I do not think Oh, sorry. What? Was somebody trying to say something? Maybe not. They're all weird trying to say something. I don't know. >> I don't think so. >> Oh, weird. Okay. Anyway, that was weird. **[18:27]** >> What' you hear? >> I don't know. It was just like somebody trying to speak. Maybe it was like Zoom stack. >> Yeah, I don't know. I I don't I didn't hear anyone say anything. >> Okay. >> But if you guys have something to comment, I mean, just interrupt her. Hype up, >> please. Like, I want I want this to be very interactive because next week you guys are going to try this yourself. Well, we're we're going to walk through some plagues. Okay. So, >> aside from Apple Intelligence 2, before we end today, let us know what types of **[18:58]** models are accessible enough so that people know what to download for next week if they don't have ondevice Apple, >> you know. >> Oh my, we have four people watching. All right, let me go check chat. We got a bunch of chat messages. Hold on, pause. Like so because like my phone isn't not a Apple Intelligence phone so I'll I'll want to know like what other stuff to try too. >> Oh trust me I'm going to get into that in just a second. We got chat messages. Hey Gaming with Ears. Welcome to the **[19:29]** channel. Thank you guys. We've had a lot of people in chat today. You guys must like the subject. Um yeah, you guys really like this. So um so anyway, yes. Uh what models? So let **[20:01]** me talk about this then. Where do you get models? Right? Cuz we're we're talking about all this. Where do you get them? There's a variety of different ways to get them. Okay. So, the ne the first way is there's a site called O Lama. O L M A.com. Let me go put that in Zoom so you guys find it. Okay. Um All right. So, lama.com. Okay. **[20:39]** me. Let's see. Hold on. Let me go put this in chat real quick. I can um hold on. Let me go put this in chat. Let me go put this in chat real fast and then we'll get started with more of these sites. So, I'd keep these in the back of your mind here **[21:11]** if I can actually chat. Okay. Hold on. Shoot. Sorry, chat's being a little strange. Okay, I don't know if I build chat. That's **[21:41]** strange. It's not letting me chat. Weird. Okay, hold on. I want to send to everybody. Is that not an option? That's strange. Uh, okay. I also forget that. So, it's O L.com. I don't know why it's not letting me uh chat. That's strange. Olama. Okay, so Olama.com. Um, this is ran by Meta aka Facebook **[22:16]** and it's really cool because it is with a lot of models. Now, something to note is that you can So, uh you're wondering how do I start with this? Now, if you're like a adventurer, a tinkerer, >> you want to get going, >> please. Oh my god. Yes. >> I'm going to give you a preview. Hold on. Let me show you. >> I'm a tinkerer. [laughter] >> All right. Hold on. Parallel, shut up. Okay. >> I don't know how many of you guys out there also love to tinker, but I am a **[22:47]** huge tinkerer. >> I'll show you, >> if that's even a word, >> I will show you what to do. So the simplest way if you want to run models in your terminal, you can do that. I'll show you. So let me uh here let me show you what I can do here. I'm going to use my Mac. >> Okay. Question question. >> Are all the ondevice AI models besides Apple, we're not counting Apple. Like all these other ones that you can download, do they run only via terminal or do they have their own apps? **[23:19]** >> No. So they >> I just want to make sure people understand the framework because I know that when you're just random searching the web or asking GPT, it can be confusing. >> So I want to start there if you don't mind. >> That's okay. So there are many different what I call model aggregators. >> Like I hope you don't mind us taking a step back for a second and talking about how these actually like run. You know what I'm saying? >> Yeah, I think that's important. >> I love talking about this stuff. >> Yeah. I just think it's important to fill that little gap in before you show **[23:50]** us real quick. >> Model aggregators. What are these? So, these are sites or different apps that you can download that will let you run multiple models for multiple different websites, multiple different providers. Okay, so these are what I keep saying is Open Web UI. It's okay. It's not very accessible. LM Studio, same thing. Not very accessible. There's also VOL Lama, but again, it's not as featurerich as some of the others. And those are the main applications. And **[24:21]** of course, they're once we get ours out, it's going to be called Perspective Studio. And I've got a huge a huge deal that I'm coming out with. So, watch out. I'm I'm getting something in the works here. So, Perspective Studio again is really cool because it will let you search models on different websites and download them and then be able to chat with them in the same application without having to do any terminal stuff. Okay, **[24:52]** so that is going to be coming out hopefully very soon and um I'm really excited about it. But besides that again there's open web UI and lm studio. Those are the main ones. And then if you want to run them in your terminal uh the main way to do that is through ol.com. Okay. That's strange. **[25:22]** >> Oh llama. I don't know. I always just think of it that way. Oh llama. >> Yeah. So, are you guys ready for something? >> Do it. Do it. >> So, first what you're going to have to do is you're going to have to install Olama. So, go to ola.com. I already have it installed. And then you can find install and install for your platform. Okay. So, that's step one. I've already done this. So, let me go find a model to pull real quick. **[25:53]** Hold on. Please wait while I think about what I'm going to download. because I got a lot on my computer already. So, let me think here. I'm going to show you guys the whole process. Nope, not doing that one. Um, let's see. Um, I will pull. So, what I'm doing right now is I'm looking through.com. **[26:26]** Let's do. And yes, there's also vision models on here, y'all. So, you could have a lot of fun putting images and replicating Be My AI. Okay, I will warn you. This this could eat up a lot of your time. Okay, just letting you know. So, let me show you guys what we do here. I'm going to turn my speech rate. Um, let's start with let's see **[27:00]** let's see me [snorts] I need to turn on my speech real quick. All right. Okay. So, I've turned it down. Let me open up a terminal. Go. **[27:32]** Are you guys ready? >> Yes. >> Here we go. >> Well, we can't hear your screen reader. >> No, you can't because I haven't shared my screen yet. >> Okay. Okay. Okay. I just didn't know if you [laughter] >> Zoom share tool window system dialog. >> All right. There we go. So, I'm doing this on a Mac. Okay. So step one find the your find the terminal application. Okay. So I'm just going to do spotlight and press >> selection deleted. T E M I L >> terminal terminal video test space **[28:02]** directory. I don't care. >> device models earlier today. >> Minimum full screen. >> Uh that didn't work too well. >> Space with terminal containing >> shut up voice over. Okay. So here's what you can do. First in order to see if you have all llama. That's step one. I'm just going to type O Lama. O L L A enter. >> CSH command not found. A llama base video test. >> Well, looks like I don't have a llama, even though I thought I did. Okay, I guess I should just show the whole process here. All right. **[28:33]** >> Yeah, you do. >> Let's do it. So, I'm going to go to my browser >> doc apps Safari terminal video test tail Safari space with applications Xcode finder and >> shut up. Open location. >> I'm going show you guys how to do this. >> N visit links menu. Download three link download visited link and end of banner heading level one down. Navigate link Mac OS link Linux link Mac OS banner **[29:04]** visited link link GitHub heading level one navigation visited link Linux heading not found. Blink end of blink download for Mac OS requires Mac OS 14. No more later. Link download for Mac download has started. >> Okay. HD collection downloads button clear button pop clear button out download for show and finder button finder screen DM open >> I just had reset my map >> now app open needs attention this item don't **[29:36]** open default finder window icon view collection one item selected >> alarm view collection one item selected real List view table row applications llama app app copy llama app now and download now and searching this Mac now English window for assignment play now paste item zoom lama open needs attention helps open default find **[30:09]** application system dialog most accessible >> Taylor password >> putting my password because I had to move it from the downloads application >> finder applications Options window list view table row 41 of 76 lama app application November 1st 2025 12 terminal terminal vid space with zoom mute spotlight spot zoom finder find safari parallels windows 11 xcode visual stoinder zoom zoom audio now muted audio **[30:39]** now unmuted spot r terminal base tilt o Hey, usage lama flags a llama command available commands serve start a llama create a model show information for a model run a model stop a running model pull pull a model from a registry push push a model to a registry sign in sign in to a llama com sign out sign out from a llama com list list models ps list running model cp copy a model rm remove a model help about any command flags h help for llama v version **[31:12]** show version information use a llama command help for more information about a command and base tilt. >> So you and by the way I have the oh my dsh that's why it reads like that. So you guys and then the base is a python environment that I always have active in case I need to mess with something. So you guys just heard Lama, right? So I went ahead downloaded it from.com. I installed it and then I typed in the terminal and guess what? I've got Olama now. So does this make sense to everyone? The same processes on Windows **[31:43]** as well >> so far. So good, I think. >> Okay. So, let me actually I need to maximize my terminal. So, >> full minimize full screen button. >> All right. >> Terminal. >> Now, I'm going to do Oh, llama. >> O L A >> space U L. >> I think it's llama l >> a m a >> llama 32 >> period 2: >> llama 3.2 2 colon latest a t latest new line pulling man ling d53s. **[32:17]** >> So what this will do is I just have this command memorized. It will pull the latest version of llama 3.2 and that's like 4 gigs. Okay. So I've got pretty fast internet. This won't take me a whole lot longer but >> 64 MBS16s. >> So you guys hear it's pulling right? So I'm actually messing with models in real time now. I'm pulling them down from lama.com and then I'm going to chat with them. Okay. So what I've done is **[32:47]** I've lama run or sorry Olama pull and if you click on a model on.com you can actually see the >> space pulling 96 pulling FCC 58 pulling a7 FF pulling 56 beep pulling 34 beep verifying shot 200 writing manifest success base tilt base tilt base tilt >> I already downloaded it >> o a n run l a m a 3 4 3 llama 3 period 2 **[33:18]** >> clock a t e s t dot 236 dots 236 >> okay >> dots 145 >> okay >> greater than greater than greater than greater than >> when you hear this greater than greater than greater than you're good to type okay so do you guys hear this right so I pulled it I'm running it now and now I can type did this all make sense again this process is on all operating systems **[33:49]** Yeah. So, whenever you said that you're pulling it from O Lama, you were downloading the model. >> Yes. Yes. >> And this is the part that would need internet connection. So, yes. And obviously I could demonstrate, but I'm not going to because I don't want to get lost out of this Zoom call. But I could theoretically turn on airplane mode and still chat with this model. Okay. But I'm not going to do that here because I **[34:20]** don't want to get disconnected from my Zoom. Uh, okay. Now I'm at the prompt. So let me just type in a question. I'm going to say what >> a what is a A S E R O >> iOS 145 of my n December 17, 2012 iOS check Apple most up to check for the slash end time for the most current information greater than greater than greater than. >> You notice this was not a good experience. Why? Cuz it's automatically streaming in the terminal and it's just **[34:52]** not a good easy like way to hear this, right? But it did work properly. It it worked. But the problem is is that this is not the best you way to read this because it's obviously like it it's just not great. So what I can do is I can just use the insert up or the V out of shell text and send a new please note new line HT new star I as of greater than greater than as of my knowledge greater than greater than greater than what is the latest version **[35:22]** of iOS. Down air uh vo down >> as of my knowledge cut off in December 2023. The latest version of iOS is >> uh oh >> new line star iOS 17.4 released on March 7th 2024. >> Interesting >> star. However, please note that Apple typically releases new versions of iOS every year and it's always best to check Apple's official website or authorized resellers for the latest and most upto-date information. New line. You can also visit Apple's website to check for **[35:52]** the latest version of iOS. New line https/sup support apple/ ns/ht2833. New line. Please note that this information may change over time and I recommend verifying with Apple's official resources for the most current information. New line greater than greater than greater than send a message slash for help. >> See, so that was all generated on device. There's no cloud activity going on there. **[36:23]** What do I mean? Yes, it's not fully accurate, but the fact is at least I could get something out of this. What do you guys think about this? >> There's other >> there's other models that are like up to date, right? >> Yeah. So, there's models that are being trained all the time. I just did this one because this is just the de facto like start here kind of thing. >> Gotcha. >> And there is support for pulling models **[36:55]** from hugging face which is another popular registry but again I don't exactly know how to do that. But let me show you. I'm going to go back to my window. Actually no. Uh I'll show you guys like you can actually look and uh see what other like models you have. So let me go back to.com. Finder Safari Safari show and finder photo http download out of image toolbar image tool down download and have banner >> visited link image lama link models **[37:27]** >> models I'll show you here we have >> GPT open weight models designed for powerful reasoning agentic tasks and versatile developer use cases tools thinking cloud 20B 120B4M pulls five tags updated 3 weeks ago link heading level two >> obvious Obviously, I can't run any of those models. >> Quen 3VL, the most powerful vision language model in the Quen model family to date. Vision Cloud 2 B4 B8B 30B 32B 235 B79.5K **[37:57]** pulls 59 tag. >> So, for example, if you got 16 gigs of RAM, you could probably get away with an 8B, which could take either 8 to 10 gigs of RAM, right? You don't want to try to run a 20B. It's just not going to work with your RAM. So that's kind of and then of course like if you got like 24 32 gigs you could probably run like a 20B or 24 like so it just kind of goes up from there but again the applications like LM Studio or other things will actually tell you if these models will work on your device or not. Okay, this will not. So that's the thing you have **[38:28]** to keep in mind with the llama.com it will not tell you. Okay, so I can let me click on one of these models real quick. Deepc one deep quen 3 coder Alib Gemma 3 the current most action menu item link heading level two page menu reader available >> it's the open source models from Google from Gemini Gemma 3 web item was removed engine navigation heading level **[39:01]** terminal zoom parallel desk windows 11 xcode visual studio Oh, Safari. Let me go back here. >> 1B, 4B. >> So, there's 4B, 1B, >> 12B, >> 12 >> 20 level two models. >> And these are all So, it says Gemma 3, which is Gemini, like open source. So, they have Gemma and Gemini. These are all Google models. >> Link view. All right. Arrow name, size, >> and this is a table. >> Contact input link. Gemma 3 latest >> latest >> button. >> That's label **[39:31]** >> 3.3GB 128K. So this basically means it's 3 GB and it's 128K context window. So when I Yeah. So that's basically what this means. >> Text image. >> Blink Gemma 3 200 button. Blink Gemma 3 270. >> I'm all about the huge context windows. >> Me too. >> Button 292 MB 32K. >> No, this one's only 32K cuz it's Remember the 70 million. Remember what I talked about? So the context window went **[40:02]** down. Text link Gemma 3 1B. >> This is a 1 billion parameter. >> Button 815 MB 32K. Text link Gemma 3 4B. Latest button 3.3GB 128K text image. >> See that one's text and image. So if you want to put images through this, you could. >> Blink Gemma 3 12B. >> This is a 12B >> button 8.1 GB. So remember, it's 8 gigs **[40:33]** for storage, but that does not mean it will take 8 gigs of RAM. It'll probably take a little bit more. >> 128K >> text image. What's a good rule from trying to figure out the difference or like when trying to figure out basically the rule of like okay I have x amount of RAM I can use up so I can afford to have x billion or whatever token not token x billion >> whatever word I'm looking for um **[41:03]** >> parameter um model running like what's a good rule of thumb to kind of calculate what the biggest one you can get away with those. >> So yeah. Um so for example, if you got a 16 um gig machine for RAM, you could probably go with with an eight or maybe a 12. 12 is pushing it. And then of course for like 32, you could go with a 20 billion parameter potentially or even >> notification from Discord left number **[41:35]** 26.1 Apple Music1. >> Okay, there we go. Those are just Apple news feeds that go through our Discord channel for important Apple news. Anyway, um so those are kind of some rules of thumb. You want to keep the like you don't want to have the same amount of RAM. >> Sure. >> For a model is what you have on your disc. >> So would it be a good idea to really like when you're not running anything to go look at your activity monitor or whatever the Windows equivalent is and sort of see like on a normal use moment **[42:06]** how much RAM you've got free. Yeah, that could be helpful. >> And then run whatever billion is just below that basically. >> Yeah. Or you know, things like that. But if you have, of course, once our application comes out, it will tell you like, hey, this you can't run this on your device or it may be slow or this will be great for your device, right? >> No, you're good. I'm just I'm just trying to like for those out there that are kind of wondering on their own how to determine what models might work on their machine, you know? I just thought it would be good to take a second and **[42:37]** talk about like how you can kind of estimate that. >> So, basically there are some general rules, right? So, if the model if you're uh six, you know, 16 gigs of RAM, which is what I have, >> I typically would want want to run eight uh eight uh B. So, 8 billion because it could take either 10 or 12 gigs of RAM depending on the model. >> So, that'd be a little bit pushing. If it's a 12B, I could maybe do it because it could take 14 gigs of RAM. That'd be **[43:08]** very close. I'd have to not have anything open. And then, of course, I can't run a 20 billion parameter, right? >> Also, some of the models when you click on them will say how many gigs of RAM they take, >> right? Okay, perfect. >> Yes. >> Yeah. I just thought it would be good to take a second and talk about that. >> Yeah. So, are are you going to be playing with this, Laurel? >> Oh, absolutely. You better believe it. So, let's try another one. And I've got uh about you know 19 minutes left. So, **[43:39]** we'll try one more and then we'll wrap it up. But and then I'll check chat. But hopefully this was really helpful. So, let's try button Gemma 3 12 button latest Gemma 3 4 B. Uh, is there an 8B >> link button latest link at third eight button link Gemma 3 1B text link but 8 third tag link latest button 3.100 text link Gemma 3 12B but 81 tech link Gemma **[44:12]** 3 20 but 71 tech heading level two read me text button blink text link Gemma 3 12B tech three button latest blink gemma 3 4B >> we'll just do this page menu reader available nav visited Heading level two, the current most vision. >> Heading level two, the current most capable model that runs on a single GPU. >> And you'll notice it has different uh like a description and then >> vision 270m **[44:42]** 1B 4B 12B 27B updated 7 months ago. A2F6 CC3F7F3.3GB dot >> okay >> link model arch gemma 3 dot 4.3b parameters >> so that's the parameters 4.3b >> 4 quantization >> and these are all things that are really advanced I'm not going to explain all of this right now >> k_m >> but cuz that's just really like advanced that'll take me about 20 30 minutes to **[45:13]** explain the quantization stuff but really what you have to know is how much it will ra run on your RAM. Quantization in short terms means that basically it kind of compresses the model almost like a zip file to try to run in smaller hardware but that could mean less quality. Okay, that's dumb down version of quantization. Okay. So, it tries to like, you know, de uh compress the model as much as it can to try to run on **[45:44]** worse hardware. But again, that can lead to worse quality. Okay, so that's just something to know. Again, we're not going to cover that right now cuz that will take me a while to explain. But let me keep going. >> 3.3 link license Gemma terms of use last modified 8 point link params left brace stop 70 link template >> template. This is for >> left brace left 300. Heading level two read me. Google Gemma 3 load heading level two the current most capable vision 20427B. Heading level two readme heading level **[46:16]** two models. Heading level three text >> text. >> Heading level three multimodal vision. Heading level three quantization a heading level two evaluation. >> There's a lot of stuff in here. And then you'll find the >> heading level three benchmark results. Heading level four reasoning logic. and heading level four multilingual capabilities. So this one might you might be interested in Laurel >> table five column benchmark column 105. >> These are what are called benchmarks. So they test these models against certain things and they publish the results. So you can look at these later but for **[46:48]** example multilingual >> GMA 3 PT1B column 2 of GEMA 3 PT4B column 3 of 5 GMA 3 PT 12B Gemma 3P2 of 8 benchmark link MGSM column 1 of 5 >> multi language >> GEMA 3 PT1B 2.04 04 column 2. [clears throat] >> I'm not going to go into all these numbers cuz they can be really confusing. >> Heading level four multimodal cape heading level two reference last heading heading level two out of toolbar image tool gem Gema 3 table benchmark tape heading out of toolbar Gemma 3 table **[47:21]** 3 PT1B 6.10 10 column met bench out of heading level four reasoning and row seven shot out of Gemma 3 heading chat heading out of toolar three benchmark table headings mod end of table out of Gemma 3 to heading level three benchmark chat heading level two out of Gemma 3 benchmark table six columns 21 rows >> me >> links menu bulk Liva big wine RX >> see I'm trying to go up >> heading level four reasoning end of t row row six **[47:52]** to benchmark row one of set 27B paratama runema 3 12B at CAD. >> So that's what you're looking for. >> 12B parl Gemma 3 4 B at CAD Gemma 3 4 B parameter model Gemma 3 4 Bit Cama Alama Gemma 3 selected 4B selected it selected >> and they have different like titles like this is how you can Oh dear. I just copied this thing. Row one to row two of seven benchmark link hell swag out of Gemma 3 27. **[48:23]** >> This site is kind of strange. Gemma 3 tape beta 3 run Gemma 3 27 cataron Gemma 3 12 cat. What we're going to do for now because I don't have much timeinal is I'll show you right you kind of got the idea of how the model kind of worked right the page I'd explore that more later but let me just show you if I wanted to pull the latest model l Gemma I'll show you that real quick >> use space greater than greater than use greater than greater than send a message slash **[48:54]** for help >> that's three greater than use plus d or slide greater than send a message base Okay. Control D to access. I'm like, what the world? Okay. Anyway, so we're going to do O lama. Run. >> O L A U P U L L G E A Gemma A >> Oops. >> A three >> Gemma 3 >> A T E Sing manifest. All need a 25 S. **[49:24]** >> I think >> 69 MB/s42. >> Yes. So, we're pulling it and then we're going to run it real quick and then we'll wrap up. I'll take chats and we'll call it a day. Okay. So, hopefully you guys have liked this. Again, um I definitely explore this more later. We'll talk more about this next week. Uh we'll just try like different applications. They've got the Olama GUI uh that just came out that I haven't really messed with yet. So, we'll talk about that. Um we'll talk about other things. But hopefully this was a good introduction to on device AI. **[49:55]** >> End of text. Pulling polling man. still working on it. >> 65 MB/S13s >> and then we'll run it. We'll ask the same question we always did and then we will go ahead and wrap up and we'll call it a day. So, let's see. Is it still >> space pulling pulling 25 pulling E08 new line pulling 3001 unpulling beast verify end of pull pulling pulling DD084 C7D92 A3% right >> I guess the last um kind of question or **[50:27]** thought I have is like so if people wanted to play with it this week with these models um how do people know what models are best for what like like >> oh that's a good question >> you know what I mean like It's cuz like you know you can do anything from interactive textbased gaming or you know fictional writing or image generation or description or coding or there's a lot of application right? **[50:57]** >> Yeah. >> How does someone have an idea what model to choose best to play around with depending on their goals. >> That's a good question. It really all depends. So there are some models it will say like code llama or code this that usually means it's good for coding right or some of them like it won't always say like creative writing or other things but that's where you could look at like the benchmarks you could kind of look online cuz it really all depends these models always change right for starter models I like Gemma because **[51:27]** it's made by Google so it's just like what we know about Gemini right so if you like Gemini you'll probably like this but it's the open source version Okay. So, it could, you know, who knows? This Gemini 3 could lead to the next version of Gemini eventually, right? So, something like that. And then, of course, the llama models are kind of the standard. They're made by uh Meta. So, if you've ever gone to Meta AI, so meta.ai, meta, whatever.com, you could actually um you know, see some **[51:58]** of those models of course as well, like on that site. And then, you know, other than that, like just play around with the models because like it all depends. Like I can't really say what's better for what, but I have tried the lava models, Lava VA. We've talked about those a little bit earlier. Those are really good vision models. So is this Gemma model as well. So playing around with that and figuring out, okay, what do you want to do? And then figuring out the best model that will run in your hardware. It can be a balancing act, but **[52:29]** definitely I recommend doing that. Ace t it's running some error pulled. So o l a m a r u r g e m m a 3 gemma 3 col a t e s t1456.14 end of 104 finder application list view table >> I want to see if I can put in a screenshot I'll describe it >> terminal >> cuz this might be list view table 41 of 76 doc so week four week four screenshot **[52:59]** 2025-10-01 at screenshot 20 terminal Yeah, let's do that. >> Space greater than greater. It's l a s e d e s c r e ais p/ users space added image apostrophe/ users/tailor/ desktop/creenshot 2025-10-01 12.48.59 48.59 p.m. PNG dots 1456 dots 124 >> okay >> dots 124 end of dot dot okay here s a **[53:30]** description quotation mark >> add mode colon leg asterisk >> PVP section light such a sync bottom L button to star and open my collect collected starst star notific learn how to star s timer star s small let me know me to zoom or interface >> it's like reading yeah wow it's it's describing this all right so Yes, it can do vision. Holy moly. >> Space greater than greater than greater than greater than greater than greater than greater than great new. Let me know if you stop new star star star **[54:02]** new star new. This screenshot shows the main menu of the Blizzard mobile game Hearthstone. It >> has a gothic inspired art style fitting with the game's theme. New line. >> Let me try that again because I interrupted it. >> The screenshot shows the main menu of the Blizzard mobile game Hearthstone. It has a distinct ornate and gothic inspired art style fitting with the game's theme. >> Okay. >> New line. Star star star. Key elements. New line. Star star star star menu tabs. Star star. There are four main tabs positioned at the top of the screen. **[54:32]** Each offering access to a different game mode. >> And this is very accurate by the way. >> Star star Hearthstone. Star star. This is the core single player campaign and quest mode. Star star battling grounds. Star star. This is the PVP battlegrounds mode. Star star star Arena. Starst Star. This is the auto battle system for quick PvP matches. Star Star Star modes. Star Star. This section likely provides access to various other modes such as single player challenges. New line. Star **[55:02]** Star Bottom Elements. Star Star Shop. Star Star A button to access the in-game store. Star Star open packs. Star Star A button to purchase and open card packs. Star star star star my collection star star star a button to view and manage your collected cards. New line star star star notification/tips bar star star in the upper right corner there's a notification bar displaying a tips message encouraging players to learn how to use the live translation feature new **[55:33]** line star star star bottom left corner star star star the game's timer is displayed 12:48 p.m. New star star star player profile icon star. There's a small icon representing the player's profile. New line. Let me know if you'd like me to. >> What do you guys think about that? All I said was describe this and that's what I got from this on device model. >> That is awesome. That's really cool. >> Yes, it is. So things like this. I'm going to go ahead and stop screen sharing because I'm running out of time. >> Zoom. I could see a really good use case **[56:05]** for that like on Mac because >> I don't have be my eyes or whatever on my Mac >> so I could I would appreciate having something like that for image description you know >> and you can access it so just do the Gemma latest so uh pull Gemma 3 colama run Gemma 3 col >> and you got to use ter your map >> terminal yes there's the Jaw's voice to read you that then **[56:36]** >> I just literally press caps lock up arrow to read it. >> I mean like how are you running the Jaw's voice on Mac? >> Oh there's a there's the eloquent voices that are already built into >> That's right. >> I don't use I'm like oh my gosh. Anyway, so it just always makes me default to running Windows on a Mac in my mind. You know what I mean? Um >> well I do have Windows on this Mac. >> Yeah. Yeah. Yeah. I know you do. So >> let me check chat real quick. >> Is in terminal. >> No chat. Okay. But yeah, there's also other gooies and things we'll play with, but hopefully you guys got a good **[57:07]** overview. I have to go assume, but any last questions? Hopefully you guys found this really cool. I did. I loved it. >> I think it's very cool. I'm going to try to figure out a different terminal app I can use or whatever because I don't want to like mess up my computer, you know. >> None of the other No. Well, >> yeah, I need to head out. So, uh, >> all right. Well, same time, same place next Tuesday. We see you guys.