WEBVTT 00:00:00.000 --> 00:00:03.819 Imagine you are on a long flight, thousands of 00:00:03.819 --> 00:00:06.719 feet in the air, you have zero Wi -Fi connection, 00:00:07.339 --> 00:00:11.099 beat, yet your iPhone is flawlessly summarizing 00:00:11.099 --> 00:00:13.599 a highly confidential contract. Yeah, completely 00:00:13.599 --> 00:00:15.660 offline. It really does feel a little bit like 00:00:15.660 --> 00:00:17.699 magic when you first see it happen. Welcome to 00:00:17.699 --> 00:00:19.940 this deep dive. I am so glad you are here with 00:00:19.940 --> 00:00:22.679 us today. Our mission is beautifully specific. 00:00:22.920 --> 00:00:25.100 Right. We are basically moving the cloud right 00:00:25.100 --> 00:00:28.379 into your pocket. Exactly. We are exploring how 00:00:28.379 --> 00:00:31.620 to turn your everyday iPhone into a highly secure, 00:00:32.020 --> 00:00:34.920 completely offline AI powerhouse. And you do 00:00:34.920 --> 00:00:38.899 this using a free app called Locally AI. We always 00:00:38.899 --> 00:00:41.539 hear that AI requires massive data centers, giant 00:00:41.539 --> 00:00:44.179 warehouses of servers humming away. They consume 00:00:44.179 --> 00:00:46.659 insane amounts of power. Yeah. That is the standard 00:00:46.659 --> 00:00:48.679 narrative. But what if that narrative is fundamentally 00:00:48.679 --> 00:00:51.990 flawed? Let us rethink the privacy paradigm shift. 00:00:52.310 --> 00:00:54.270 This is where the underlying mechanics get really 00:00:54.270 --> 00:00:56.750 fascinating. Think about how most AI operates 00:00:56.750 --> 00:00:59.810 today. It usually asks you to log in with Google 00:00:59.810 --> 00:01:03.009 or Apple. You instantly create a digital footprint. 00:01:03.590 --> 00:01:05.849 You are linked to your real name. Online apps 00:01:05.849 --> 00:01:08.989 process your prompts on servers you cannot see. 00:01:09.950 --> 00:01:13.810 And they often log that data to train future 00:01:13.810 --> 00:01:16.890 models. It creates a tremendous amount of anxiety 00:01:16.890 --> 00:01:20.900 for people. You worry about feeding sensitive 00:01:20.900 --> 00:01:24.019 information into a corporate server. Yeah, absolutely. 00:01:24.040 --> 00:01:27.019 Like a corporate strategy or personal journaling. 00:01:27.379 --> 00:01:29.680 You just do not know who has access to it. Right. 00:01:29.799 --> 00:01:32.219 And that is precisely why cutting the cord is 00:01:32.219 --> 00:01:35.939 so revolutionary. An offline AI app fundamentally 00:01:35.939 --> 00:01:38.540 changes the physics of data privacy. Because 00:01:38.540 --> 00:01:41.040 there is no connection. Exactly. Locally, AI 00:01:41.040 --> 00:01:43.980 requires zero login. No name, no email address, 00:01:44.000 --> 00:01:46.239 no phone number. Wow. Because the app does not 00:01:46.239 --> 00:01:48.640 need to communicate with a server. It simply 00:01:48.640 --> 00:01:51.099 does not ask for your credentials. Every single 00:01:51.099 --> 00:01:52.920 calculation stays trapped inside your iPhone. 00:01:53.200 --> 00:01:55.939 No data goes out. It physically cuts off the 00:01:55.939 --> 00:01:58.099 data supply to big tech companies. Right. And 00:01:58.099 --> 00:02:00.219 that happens because all the processing runs 00:02:00.219 --> 00:02:02.579 locally. It relies entirely on the Apple Silicon 00:02:02.569 --> 00:02:05.329 inside your device. It acts just like a super 00:02:05.329 --> 00:02:07.989 smart calculator. Exactly. When you delete a 00:02:07.989 --> 00:02:10.930 chat, the memory is wiped permanently. There 00:02:10.930 --> 00:02:13.349 is no backup server holding on to your deleted 00:02:13.349 --> 00:02:16.530 questions. Using a cloud AI is sort of like speaking 00:02:16.530 --> 00:02:18.669 your thoughts through a public megaphone. You 00:02:18.669 --> 00:02:20.490 hope people are not recording it. But they could 00:02:20.490 --> 00:02:23.050 be. Right. Offline AI is like working inside 00:02:23.050 --> 00:02:27.789 a locked soundproof vault beat. Does going offline 00:02:27.789 --> 00:02:30.930 actually stop Big Tech from tracking this specific 00:02:30.930 --> 00:02:34.210 data entirely? Absolutely. If your phone antennas 00:02:34.210 --> 00:02:37.550 are not transmitting to a server, the physical 00:02:37.550 --> 00:02:40.650 pathways for that data simply do not exist. There 00:02:40.650 --> 00:02:43.310 are no hidden trackers phoning home. So zero 00:02:43.310 --> 00:02:45.889 internet means absolutely zero data harvesting. 00:02:46.349 --> 00:02:48.610 Period. That is the beauty of it. Which brings 00:02:48.610 --> 00:02:51.389 up a really profound mechanical problem. If the 00:02:51.389 --> 00:02:53.750 data is not leaving the phone, the entire brain 00:02:53.750 --> 00:02:56.330 has to live inside the device. Yeah, it sounds 00:02:56.330 --> 00:02:58.870 impossible at first. How on earth do we fit a 00:02:58.870 --> 00:03:01.680 massive AI into a piece of glass and metal? Let 00:03:01.680 --> 00:03:03.759 us look at sizing the brain in the setup. We 00:03:03.759 --> 00:03:05.800 have to understand how we actually measure the 00:03:05.800 --> 00:03:08.340 size of an AI. We categorize these models by 00:03:08.340 --> 00:03:10.840 number ending with a capital B. And the B stands 00:03:10.840 --> 00:03:13.699 for billion parameters. Right. Parameters are 00:03:13.699 --> 00:03:16.240 the digital brain cells making up the AI. More 00:03:16.240 --> 00:03:19.419 parameters mean the AI has a denser web of knowledge. 00:03:19.580 --> 00:03:23.099 It is generally smarter, but it requires a significantly 00:03:23.099 --> 00:03:26.360 larger workspace to function. Yeah. And in a 00:03:26.360 --> 00:03:28.539 phone, that workspace is your RAM. It is your 00:03:28.539 --> 00:03:30.879 phone's short -term memory. You have to match 00:03:30.879 --> 00:03:34.020 the AI brain size to the physical RAM inside 00:03:34.020 --> 00:03:36.780 your specific hardware. So for older iPhones, 00:03:36.780 --> 00:03:40.060 you use the smaller model. Exactly. You use 0 00:03:40.060 --> 00:03:44.099 .8 B models like small M. They are very lightweight 00:03:44.099 --> 00:03:47.979 and run incredibly fast. Then you step up to 00:03:47.979 --> 00:03:51.319 the two billion parameter models. The two B models 00:03:51.319 --> 00:03:53.740 like Quinn 2 .5 are really the current sweet 00:03:53.740 --> 00:03:55.800 spot. Right. They are smart enough to handle 00:03:55.800 --> 00:03:59.020 about 80 % of daily tasks, but they stay stable 00:03:59.020 --> 00:04:01.120 without crashing your phone. But what if I have 00:04:01.120 --> 00:04:03.319 one of the newest devices? Can I push it further? 00:04:03.530 --> 00:04:07.490 You can. The 4B to 8B models like Lama 3 or 3 00:04:07.490 --> 00:04:10.069 .2 are massive. Those are really only viable 00:04:10.069 --> 00:04:13.650 on the iPhone 15 Pro or the 16. Those newer phones 00:04:13.650 --> 00:04:16.129 finally have enough internal RAM to hold those 00:04:16.129 --> 00:04:18.209 heavier brains open. It is a lot of data. It 00:04:18.209 --> 00:04:22.649 is. Bet, whoa, imagine fitting two billion parameters 00:04:22.649 --> 00:04:25.129 right in your pocket. You literally carry a localized 00:04:25.129 --> 00:04:28.230 brain in your jeans. It is genuinely a marvel 00:04:28.230 --> 00:04:33.490 of modern engineering. Two secs silence. So if 00:04:33.490 --> 00:04:35.350 I am holding my phone right now trying to actually 00:04:35.350 --> 00:04:37.930 make this happen What is the physical bridge 00:04:37.930 --> 00:04:39.949 between reading about this and actually running 00:04:39.949 --> 00:04:43.509 it? It takes maybe three minutes first grab your 00:04:43.509 --> 00:04:45.430 iPhone and connect to a strong Wi -Fi signal 00:04:45.560 --> 00:04:47.800 You only need the internet this one time for 00:04:47.800 --> 00:04:50.779 the initial download. Right. You search for locally 00:04:50.779 --> 00:04:53.920 AI in the App Store. It is an Apple -checked 00:04:53.920 --> 00:04:56.980 app, completely clean, zero ads. Once you open 00:04:56.980 --> 00:04:58.879 it, I assume you have to choose which brain to 00:04:58.879 --> 00:05:00.579 install. Right. You will see a list of different 00:05:00.579 --> 00:05:03.120 models. You tap the clad icon next to your chosen 00:05:03.120 --> 00:05:05.439 brain. And how big is that file? It downloads 00:05:05.439 --> 00:05:07.519 a file that is roughly one to two gigabytes. 00:05:07.899 --> 00:05:10.000 You will want to keep about five gigabytes free 00:05:10.000 --> 00:05:12.629 on your phone overall just to be safe. And it 00:05:12.629 --> 00:05:15.370 is probably best not to switch to other apps 00:05:15.370 --> 00:05:18.009 during this download. Let the file system do 00:05:18.009 --> 00:05:21.829 its job. Exactly. Once it finishes, the fun begins. 00:05:22.490 --> 00:05:25.009 You turn off your Wi -Fi entirely, switch your 00:05:25.009 --> 00:05:27.930 phone into airplane mode, then just type hello 00:05:27.930 --> 00:05:30.449 in the chat box. If it answers you right away, 00:05:30.670 --> 00:05:33.470 congratulations. You have a working, totally 00:05:33.470 --> 00:05:37.160 offline AI. That is amazing. Yeah. And advanced 00:05:37.160 --> 00:05:40.800 users can even sideload custom .gguf files. Okay. 00:05:40.819 --> 00:05:43.420 ggus is a file format for running compressed 00:05:43.420 --> 00:05:46.079 AI locally. You pull those from sites like Hugging 00:05:46.079 --> 00:05:48.800 Face. Why shouldn't I just force an 8 billion 00:05:48.800 --> 00:05:51.959 parameter model onto my base iPhone to get the 00:05:51.959 --> 00:05:54.060 smartest answers? Because the RAM bottleneck 00:05:54.060 --> 00:05:56.480 will choke the system entirely. The AI will take 00:05:56.480 --> 00:05:59.199 a painfully long time to process the math. Right. 00:05:59.370 --> 00:06:01.829 Your phone will violently overheat, and the app 00:06:01.829 --> 00:06:04.490 will inevitably crash. Got it. A brain too big 00:06:04.490 --> 00:06:07.410 just causes lag, overheating, and massive frustration. 00:06:07.670 --> 00:06:09.529 You have to respect the hardware limits. Now 00:06:09.529 --> 00:06:11.769 that we have a smaller, highly compressed brain 00:06:11.769 --> 00:06:14.050 installed, the interaction changes. It really 00:06:14.050 --> 00:06:15.569 does. We have to talk to it differently than 00:06:15.569 --> 00:06:18.529 we do with massive cloud servers. Let us explore 00:06:18.529 --> 00:06:21.069 the fine art of offline prompting. This is where 00:06:21.069 --> 00:06:24.639 people usually stumble. You cannot treat a 2B 00:06:24.639 --> 00:06:27.699 model like the giant online versions. Right. 00:06:28.240 --> 00:06:31.259 Because the localized brain is smaller, it lacks 00:06:31.259 --> 00:06:34.120 the vast contextual safety nets of a massive 00:06:34.120 --> 00:06:36.920 server. You have to be exceptionally clear. You 00:06:36.920 --> 00:06:39.139 have to guide it. Think of the offline AI as 00:06:39.139 --> 00:06:42.000 a brilliant but easily distracted assistant. 00:06:42.259 --> 00:06:44.189 That is a great way to frame it. You cannot just 00:06:44.189 --> 00:06:46.529 type, you know, write an email to my boss. You 00:06:46.529 --> 00:06:48.970 have to explain who you are, what the tone should 00:06:48.970 --> 00:06:51.610 be, and exactly what the email must achieve. 00:06:52.449 --> 00:06:54.689 Smaller brains need incredibly strict boundaries. 00:06:54.910 --> 00:06:57.709 Otherwise, they start making things up. I still 00:06:57.709 --> 00:06:59.350 wrestle with getting the prompts right myself, 00:06:59.490 --> 00:07:01.910 even on the massive cloud models. It is a delicate 00:07:01.910 --> 00:07:04.089 balancing act. Oh, we all do. You want to give 00:07:04.089 --> 00:07:06.189 enough detail without overcomplicating it. It 00:07:06.189 --> 00:07:08.610 is a learned skill. The source material offers 00:07:08.610 --> 00:07:11.050 a brilliant template specifically for these smaller 00:07:11.050 --> 00:07:14.649 offline models. First, you assign the AI a highly 00:07:14.649 --> 00:07:17.629 specific role. You tell it. Act as a professional 00:07:17.629 --> 00:07:20.310 tutor with 20 years of experience. By assigning 00:07:20.310 --> 00:07:22.470 a role, you are narrowing down the neural pathways 00:07:22.470 --> 00:07:24.509 it needs to search. Right. It filters out the 00:07:24.509 --> 00:07:27.389 noise. Next, you state exactly what you want 00:07:27.389 --> 00:07:30.350 to learn. Then, and this is crucial, you heavily 00:07:30.350 --> 00:07:32.430 limit the length of the response. You keep it 00:07:32.430 --> 00:07:35.670 brief. Yes. You might say, explain the main idea 00:07:35.670 --> 00:07:38.470 in three very simple sentences. You can also 00:07:38.470 --> 00:07:40.649 ask for three interesting facts that most people 00:07:40.649 --> 00:07:43.269 do not know just to keep it engaging. Making 00:07:43.269 --> 00:07:45.730 it interactive is the final piece of the puzzle. 00:07:45.800 --> 00:07:49.240 Ask it to generate a short quiz with two questions. 00:07:50.240 --> 00:07:53.459 This format works beautifully. It tells the AI 00:07:53.459 --> 00:07:56.560 exactly who to be, prevents it from rambling, 00:07:56.959 --> 00:08:00.079 and forces an interactive dialogue. Why are these 00:08:00.079 --> 00:08:02.100 strict boundaries so much more critical for a 00:08:02.100 --> 00:08:05.439 2B model? Because a 2 billion parameter model 00:08:05.439 --> 00:08:07.819 simply does not have the processing power to 00:08:07.819 --> 00:08:11.040 filter out irrelevant context on its own. If 00:08:11.040 --> 00:08:13.560 you leave a prompt open -ended, it loses the 00:08:13.560 --> 00:08:16.300 thread and wanders completely off track. Constraints 00:08:16.300 --> 00:08:19.120 keep the smaller AI focused so it doesn't ramble 00:08:19.120 --> 00:08:22.629 or get confused. Sponsor. We've mastered the 00:08:22.629 --> 00:08:25.209 text prompts. But what if this localized brain 00:08:25.209 --> 00:08:27.550 could actually see and hear its surroundings? 00:08:27.870 --> 00:08:30.009 Now we're getting into the really futuristic 00:08:30.009 --> 00:08:33.240 stuff. And what happens when we inevitably push 00:08:33.240 --> 00:08:36.440 its smaller capacity just a bit too far? Let 00:08:36.440 --> 00:08:39.919 us discuss senses, silly mistakes, and hallucinations. 00:08:40.320 --> 00:08:43.299 Adding senses is arguably the most thrilling 00:08:43.299 --> 00:08:46.279 part of locally AI. It is not just a simple text 00:08:46.279 --> 00:08:48.860 box. Right. If you choose to download one of 00:08:48.860 --> 00:08:51.419 the specific vision models, your app can suddenly 00:08:51.419 --> 00:08:54.240 understand photographs. Imagine you are walking 00:08:54.240 --> 00:08:57.320 down a street in Tokyo. You see a menu or transit 00:08:57.320 --> 00:08:59.480 sign in a foreign language? Yeah, and you have 00:08:59.480 --> 00:09:01.620 zero international cell service. You just take 00:09:01.620 --> 00:09:03.440 a photo through the app and ask, what does the 00:09:03.440 --> 00:09:06.820 sign say? Wow. It processes the pixels, translates 00:09:06.820 --> 00:09:08.899 the text, and gives you the answer instantly. 00:09:09.179 --> 00:09:11.899 It does not use a single drop of cellular data. 00:09:12.110 --> 00:09:14.350 You can take a photo of scattered computer parts 00:09:14.350 --> 00:09:17.210 to get build advice. You can snap a picture of 00:09:17.210 --> 00:09:19.309 the ingredients inside your fridge to get instant 00:09:19.309 --> 00:09:21.850 recipe ideas. Or take a photo of a computer monitor 00:09:21.850 --> 00:09:24.710 to find mistakes in your code. It is wildly versatile. 00:09:24.870 --> 00:09:27.049 And for voice interaction, it is just as seamless. 00:09:27.330 --> 00:09:29.529 Exactly. You hold the microphone icon and speak 00:09:29.529 --> 00:09:31.950 your prompt. You just have to download a tiny 00:09:31.950 --> 00:09:35.129 voice processing file beforehand. Talking to 00:09:35.129 --> 00:09:38.450 an AI assistant in the middle of a remote off 00:09:38.450 --> 00:09:41.450 the grid forest is quite a surreal experience. 00:09:41.710 --> 00:09:45.970 beat. However, we must ground this in reality. 00:09:46.710 --> 00:09:49.789 It is not invaluable. No, it definitely is not. 00:09:49.830 --> 00:09:52.549 Sometimes it fails spectacularly. In the tech 00:09:52.549 --> 00:09:55.210 world, we call this a hallucination. Right. Hallucinations 00:09:55.210 --> 00:09:57.710 happen when the AI confidently invents facts 00:09:57.710 --> 00:10:00.590 it simply does not know. Because the AI is designed 00:10:00.590 --> 00:10:03.769 to predict the next logical word. it guesses 00:10:03.769 --> 00:10:06.549 when it lacks the actual factual anchor. Remember, 00:10:06.850 --> 00:10:09.649 this AI was heavily compressed to fit inside 00:10:09.649 --> 00:10:12.450 a phone. It is like a scholar who read thousands 00:10:12.450 --> 00:10:14.730 of books but completely lost the index. That 00:10:14.730 --> 00:10:16.850 is a great analogy. It remembers the general 00:10:16.850 --> 00:10:19.690 concepts, but it forgets which page a specific 00:10:19.690 --> 00:10:22.289 fact was printed on. It frequently fails at hard 00:10:22.289 --> 00:10:25.080 math. It struggles with highly specific historical 00:10:25.080 --> 00:10:28.240 dates. And it obviously cannot tell you today's 00:10:28.240 --> 00:10:30.200 news because it has no internet connection to 00:10:30.200 --> 00:10:32.700 check. There is a famous benchmark test right 00:10:32.700 --> 00:10:35.940 now. You ask the AI how many letters are in the 00:10:35.940 --> 00:10:38.840 word strawberry. Oh, I have seen this. The offline 00:10:38.840 --> 00:10:41.320 models will sometimes confidently look you in 00:10:41.320 --> 00:10:43.860 the digital eye and say there are only two. It 00:10:43.860 --> 00:10:47.120 struggles with spelling because AI processes 00:10:47.120 --> 00:10:50.350 text in chunks. Tokens are just the small text 00:10:50.350 --> 00:10:53.190 chunks in AI processes. It does not see individual 00:10:53.190 --> 00:10:56.730 letters. This is why we must treat it as a helpful 00:10:56.730 --> 00:10:59.669 assistant, not a perfect guide. If it's guessing 00:10:59.669 --> 00:11:02.169 facts, is it really safe to use for learning? 00:11:02.429 --> 00:11:05.169 It is perfectly safe if you understand the limitations. 00:11:05.570 --> 00:11:07.870 You use it for creative work, summarizing long 00:11:07.870 --> 00:11:10.950 documents, or brainstorming broad concepts. Right. 00:11:11.049 --> 00:11:13.529 But you must actively double -check important 00:11:13.529 --> 00:11:16.169 facts, like precise historical dates or specific 00:11:16.169 --> 00:11:18.809 medical claims. Use it for concepts and creativity, 00:11:19.250 --> 00:11:21.590 but double -check hard facts and dates. Exactly. 00:11:21.870 --> 00:11:24.820 Trust would verify. Always. Making a phone think 00:11:24.820 --> 00:11:28.039 this hard, visually and verbally, has intense 00:11:28.039 --> 00:11:30.279 physical consequences on the hardware. It really 00:11:30.279 --> 00:11:32.799 pushes the limits. Let us examine the real world 00:11:32.799 --> 00:11:35.679 physics of running offline AI. We need to explore 00:11:35.679 --> 00:11:38.860 battery life, heat generation, and RAM management. 00:11:39.049 --> 00:11:41.950 Running an AI model locally is arguably the most 00:11:41.950 --> 00:11:44.529 demanding task you can ask a phone to do. It 00:11:44.529 --> 00:11:46.870 is functionally similar to playing a massive 00:11:46.870 --> 00:11:50.149 heavy graphical game like Genshin Impact. The 00:11:50.149 --> 00:11:53.570 internal Apple chip has to perform billions of 00:11:53.570 --> 00:11:55.889 matrix multiplication operations every single 00:11:55.889 --> 00:11:58.629 second. It is pulling five to ten times more 00:11:58.629 --> 00:12:01.379 power than casually browsing a website. Which 00:12:01.379 --> 00:12:04.259 means your phone is going to get hot. That thermal 00:12:04.259 --> 00:12:06.600 output is entirely normal, especially if you 00:12:06.600 --> 00:12:10.399 are running a larger 4B or 8B model. But we have 00:12:10.399 --> 00:12:12.519 to actively manage the phone's memory to keep 00:12:12.519 --> 00:12:15.259 it stable. RAM is the short -term memory your 00:12:15.259 --> 00:12:18.139 phone uses to juggle active tasks. If the desk 00:12:18.139 --> 00:12:21.159 is cluttered, the AI has no room to work. Right. 00:12:21.259 --> 00:12:24.360 An offline AI app absolutely needs space to breathe. 00:12:24.440 --> 00:12:26.600 Before you start, you must close other heavy 00:12:26.600 --> 00:12:30.059 apps. If you have 50 browser tabs open and a 00:12:30.059 --> 00:12:32.519 video game suspended in the background, the AI 00:12:32.519 --> 00:12:34.960 will simply not have enough RAM to load its parameter. 00:12:35.159 --> 00:12:37.019 Clearing out that background clutter makes the 00:12:37.019 --> 00:12:39.820 AI respond significantly faster. You also have 00:12:39.820 --> 00:12:42.240 to manage the AI's own memory. You need to clear 00:12:42.240 --> 00:12:45.340 out your old AI chats. Yes. The context window 00:12:45.340 --> 00:12:47.899 takes up RAM. Context window is everything you 00:12:47.899 --> 00:12:50.730 have typed in one conversation. If the AI has 00:12:50.730 --> 00:12:54.090 to remember too much chat history, it slows to 00:12:54.090 --> 00:12:57.230 an absolute crawl. The best practice is to start 00:12:57.230 --> 00:12:59.850 a brand new chat whenever you switch to a totally 00:12:59.850 --> 00:13:02.110 new topic. We also need to protect the physical 00:13:02.110 --> 00:13:04.610 battery health. You really should not use the 00:13:04.610 --> 00:13:07.090 AI heavily while the phone is actively plugged 00:13:07.090 --> 00:13:09.629 into a charger. Charging a battery generates 00:13:09.629 --> 00:13:12.769 heat. Processing AI generates heat. Combining 00:13:12.769 --> 00:13:15.740 them makes the phone dangerously hot. Yeah. You 00:13:15.740 --> 00:13:17.759 should also try to keep your battery level above 00:13:17.759 --> 00:13:22.039 20%. iOS will automatically throttle performance 00:13:22.039 --> 00:13:25.080 to save power if your battery gets too low, making 00:13:25.080 --> 00:13:27.980 the AI sluggish. It is like asking someone to 00:13:27.980 --> 00:13:30.220 run a marathon while wearing a thick winter coat. 00:13:30.779 --> 00:13:33.080 You need to take off heavy phone cases so the 00:13:33.080 --> 00:13:35.159 internal processor can actually breathe. Right, 00:13:35.340 --> 00:13:37.720 and just like that winter coat, a thick rubber 00:13:37.720 --> 00:13:40.159 case traps the thermal energy right against the 00:13:40.159 --> 00:13:43.440 glass. The phone cannot shed the heat. Can something 00:13:43.440 --> 00:13:45.879 as simple as a phone case actually slow down 00:13:45.879 --> 00:13:49.039 the AI's answers? Absolutely. Apple silicon chips 00:13:49.039 --> 00:13:51.440 have strict thermal limits to prevent permanent 00:13:51.440 --> 00:13:54.639 hardware damage. If the heat cannot escape, the 00:13:54.639 --> 00:13:57.700 chip deliberately throttles its speed. It mathematically 00:13:57.700 --> 00:14:00.279 slows itself down to cool off. Yes, trapped heat 00:14:00.279 --> 00:14:03.460 forces the Apple chip to throttle and slow everything 00:14:03.460 --> 00:14:05.919 down. You have to let the hardware dissipate 00:14:05.919 --> 00:14:08.320 that energy. Given these strict hardware limits, 00:14:08.580 --> 00:14:11.580 the battery drain, and the occasional hallucinations, 00:14:12.519 --> 00:14:15.919 When should we actually use this? How does it 00:14:15.919 --> 00:14:19.779 stack up against massive online models? Let us 00:14:19.779 --> 00:14:22.620 look at the big comparison between offline AI 00:14:22.620 --> 00:14:25.539 and cloud AI. This is the question people wrestle 00:14:25.539 --> 00:14:28.200 with constantly. The truth is, neither one is 00:14:28.200 --> 00:14:30.559 objectively superior in a vacuum. Right. One 00:14:30.559 --> 00:14:33.340 is just more appropriate for your specific environment 00:14:33.340 --> 00:14:36.620 at that exact moment. ChatGPT is like a giant 00:14:36.620 --> 00:14:40.480 comprehensive library downtown. Locally, AI is 00:14:40.480 --> 00:14:42.980 like a private smart notebook that you keep in 00:14:42.980 --> 00:14:45.259 your jacket pocket. Exactly. If you are at a 00:14:45.259 --> 00:14:47.919 desk and you need perfect coding accuracy, complex 00:14:47.919 --> 00:14:50.500 logical reasoning, and live news updates, the 00:14:50.500 --> 00:14:54.039 giant library wins easily. Yeah. NPT requires 00:14:54.039 --> 00:14:56.500 a constant data connection, and the advanced 00:14:56.500 --> 00:15:02.820 versions cost money every single month. It shines 00:15:02.820 --> 00:15:08.100 in totally different scenarios. I love using 00:15:08.100 --> 00:15:11.720 it to practice speaking foreign languages. You 00:15:11.720 --> 00:15:14.279 can prompt the offline AI to act like a native 00:15:14.279 --> 00:15:18.029 speaker from Madrid. You can have a fluid spoken 00:15:18.029 --> 00:15:20.669 conversation in Spanish while walking through 00:15:20.669 --> 00:15:24.070 a remote state park. You are practicing a complex 00:15:24.070 --> 00:15:26.889 new skill without pinging a single cell tower. 00:15:27.049 --> 00:15:29.190 It gives you high quality learning completely 00:15:29.190 --> 00:15:32.070 off the grid. It removes the latency of waiting 00:15:32.070 --> 00:15:34.710 for a server to respond. It just makes daily 00:15:34.710 --> 00:15:36.889 life much more interesting. I think the most 00:15:36.889 --> 00:15:39.610 rational approach is a hybrid workflow. You sit 00:15:39.610 --> 00:15:41.529 at your desk with a strong Wi -Fi connection 00:15:41.529 --> 00:15:45.090 and use cloud AI for heavy, complex analytical. 00:15:45.000 --> 00:15:48.600 lifting. Then you use the offline AI out in the 00:15:48.600 --> 00:15:51.720 real world for brainstorming, quick translations, 00:15:52.059 --> 00:15:55.639 and absolute privacy. So the goal isn't to replace 00:15:55.639 --> 00:15:59.059 cloud AI entirely? Not at all. It is about expanding 00:15:59.059 --> 00:16:01.620 your personal toolkit. You choose the massive 00:16:01.620 --> 00:16:04.399 server for raw computational power, and you choose 00:16:04.399 --> 00:16:07.159 the local chip for privacy and portability. Exactly. 00:16:07.500 --> 00:16:09.559 It's a specialized tool for absolute privacy 00:16:09.559 --> 00:16:12.000 and off -the -grid access. That is the perfect 00:16:12.000 --> 00:16:14.120 way to look at it. This brings us to our big 00:16:14.120 --> 00:16:17.879 idea recap. We have explored a truly profound 00:16:17.879 --> 00:16:20.799 shift in how we interact with computational intelligence. 00:16:21.820 --> 00:16:24.000 We are empowering our personal devices to think 00:16:24.000 --> 00:16:27.039 locally and independently. We are actively taking 00:16:27.039 --> 00:16:30.000 back control of our own personal data. We are 00:16:30.000 --> 00:16:32.679 achieving a new kind of digital freedom where 00:16:32.679 --> 00:16:35.679 you do not have to sacrifice technological utility 00:16:35.679 --> 00:16:38.360 to maintain your privacy. You no longer have 00:16:38.360 --> 00:16:41.840 to depend on distant opaque server forms to process 00:16:41.840 --> 00:16:44.159 your thoughts. You can carry a highly capable 00:16:44.159 --> 00:16:47.419 mind right inside your pocket, completely severed 00:16:47.419 --> 00:16:49.519 from the surveillance economy. And by doing that, 00:16:49.659 --> 00:16:52.039 you cut off the constant data harvesting that 00:16:52.039 --> 00:16:54.379 fuels big tech companies. You become a user again, 00:16:54.559 --> 00:16:57.039 rather than a product. It leaves us with a fascinating, 00:16:57.299 --> 00:16:59.720 almost unsettling question to consider. Two -sec 00:16:59.720 --> 00:17:02.940 silence. If our pocket devices can now hold highly 00:17:02.940 --> 00:17:05.880 capable, private AI brains that don't need the 00:17:05.880 --> 00:17:08.730 Webber, How will this change our fundamental 00:17:08.730 --> 00:17:10.589 relationship with the internet in five years? 00:17:11.009 --> 00:17:13.390 Will going offline become the ultimate luxury? 00:17:13.569 --> 00:17:15.890 That is a truly profound thought to end on. It 00:17:15.890 --> 00:17:18.069 changes everything about how we value connection. 00:17:18.569 --> 00:17:20.269 Thank you so much for joining us on this deep 00:17:20.269 --> 00:17:22.230 dive. I highly encourage you to download the 00:17:22.230 --> 00:17:24.670 app and test the airplane mode trick yourself. 00:17:25.190 --> 00:17:27.230 Try talking to on your next commute or your next 00:17:27.230 --> 00:17:29.829 walk in the park. Stay curious, protect your 00:17:29.829 --> 00:17:33.109 privacy, and we will see you next time. ODTRO 00:17:33.109 --> 00:17:33.529 music.