WEBVTT 00:00:00.000 --> 00:00:03.100 Most people have quietly accepted a very strange 00:00:03.100 --> 00:00:08.080 compromise. Beat. You pay $20 every month for 00:00:08.080 --> 00:00:11.160 an AI that monitors you. Yeah, it actively tracks 00:00:11.160 --> 00:00:13.599 your daily work. It routinely slows down during 00:00:13.599 --> 00:00:15.679 peak hours, too. And it sends your sensitive 00:00:15.679 --> 00:00:17.620 thoughts straight to a corporate server. Right, 00:00:17.719 --> 00:00:20.800 but Google's release of Gemma 4 shattered that 00:00:20.800 --> 00:00:23.739 dynamic entirely. Absolutely. The era of renting 00:00:23.739 --> 00:00:26.320 intelligence from the cloud is officially over. 00:00:26.570 --> 00:00:28.989 Welcome to the deep dive. We have a lot of ground 00:00:28.989 --> 00:00:32.149 to cover today. Today we are analyzing the reality 00:00:32.149 --> 00:00:35.729 of Google Gemma 4. We are looking at why running 00:00:35.729 --> 00:00:39.009 a massive model locally changes everything. And 00:00:39.009 --> 00:00:40.630 we will break down the mechanics of installing 00:00:40.630 --> 00:00:43.130 it yourself. It really is about true data sovereignty. 00:00:43.310 --> 00:00:45.030 Yeah, it is a free model you run on your own 00:00:45.030 --> 00:00:47.609 hardware. Let's unpack the foundational philosophy 00:00:47.609 --> 00:00:51.009 behind this shift first. Why does moving computation 00:00:51.009 --> 00:00:53.909 to your desk actually matter? Because it completely 00:00:53.909 --> 00:00:55.810 eliminates the middleman from your workflow. 00:00:56.200 --> 00:00:58.740 You are no longer relying on a constant internet 00:00:58.740 --> 00:01:01.460 connection. Or paying those endless monthly subscription 00:01:01.460 --> 00:01:04.840 fees. Exactly. When you use cloud tools, your 00:01:04.840 --> 00:01:07.920 data is the product. Your proprietary code or 00:01:07.920 --> 00:01:10.379 your financial spreadsheets get ingested by their 00:01:10.379 --> 00:01:13.280 servers. But Gemma 4 runs purely on your local 00:01:13.280 --> 00:01:16.819 silicon. That represents a massive philosophical 00:01:16.819 --> 00:01:19.079 pivot for personal computing. It really does. 00:01:19.180 --> 00:01:21.359 You aren't just... querying an Oracle anymore. 00:01:21.599 --> 00:01:23.799 You literally own the Oracle. Right, and you 00:01:23.799 --> 00:01:26.540 are finally liberated from infrastructure bottlenecks. 00:01:26.819 --> 00:01:30.060 When the globe logs on at 9 a .m., cloud models 00:01:30.060 --> 00:01:33.400 throttle you. With Gemma 4, your speed is dictated 00:01:33.400 --> 00:01:36.239 only by your hardware. It also features incredibly 00:01:36.239 --> 00:01:38.780 robust multimodal capabilities. Yeah, you can 00:01:38.780 --> 00:01:41.379 feed it text, images, and raw audio directly, 00:01:41.500 --> 00:01:43.799 and now you can drop a complex text document 00:01:43.799 --> 00:01:46.560 right into the chat. and ask it to extract specific 00:01:46.560 --> 00:01:49.299 line items. Or explain really anomalous data 00:01:49.299 --> 00:01:53.140 trends. Two -sec silence. Whoa. Imagine having 00:01:53.140 --> 00:01:55.739 that kind of raw reasoning power totally offline. 00:01:56.420 --> 00:01:58.459 It fundamentally redefines what a personal computer 00:01:58.459 --> 00:02:00.480 is capable of doing. It absolutely turns your 00:02:00.480 --> 00:02:02.620 machine into a secure reasoning engine. You never 00:02:02.620 --> 00:02:04.959 have to worry about API usage bills again. If 00:02:04.959 --> 00:02:07.060 it is free, what stops Google from collecting 00:02:07.060 --> 00:02:09.840 the data anyway? Well, it physically runs on 00:02:09.840 --> 00:02:12.300 your hard drive. The data never actually leaves 00:02:12.300 --> 00:02:14.860 your machine. Right. So your private data genuinely 00:02:14.860 --> 00:02:17.259 never leaves your laptop. Exactly. But you can't 00:02:17.259 --> 00:02:19.740 just cram a massive model onto an old laptop. 00:02:20.099 --> 00:02:22.539 The critical choke point is your machine's active 00:02:22.539 --> 00:02:25.539 memory. Think of RAM as your computer's short 00:02:25.539 --> 00:02:28.060 -term memory. Yeah. Get the match wrong, and 00:02:28.060 --> 00:02:31.280 it barely moves. The system has to load billions 00:02:31.280 --> 00:02:33.860 of mathematical weights into RAM. If you lack 00:02:33.860 --> 00:02:37.360 memory, the entire system grinds to a halt. That 00:02:37.360 --> 00:02:40.259 is why Gemma 4 comes in four distinct sizes. 00:02:40.860 --> 00:02:43.020 Right. The smallest versions are the E2B and 00:02:43.020 --> 00:02:46.319 E4B models. The E2B model operates smoothly on 00:02:46.319 --> 00:02:49.159 just five gigabytes of memory. And the E4B requires 00:02:49.159 --> 00:02:52.360 roughly eight gigabytes of RAM. That one is the 00:02:52.360 --> 00:02:54.319 recommended starting point for most users. It 00:02:54.319 --> 00:02:56.819 is the absolute sweet spot for balancing logic 00:02:56.819 --> 00:02:59.460 and hardware. It fully handles text, images, 00:02:59.780 --> 00:03:01.879 and audio processing. Then we cross into the 00:03:01.879 --> 00:03:04.599 heavier compute tier with the 26B model. This 00:03:04.599 --> 00:03:07.659 requires 16 to 20 gigabytes of unified memory 00:03:07.659 --> 00:03:10.879 to run properly, but it utilizes a complex mixture 00:03:10.879 --> 00:03:12.800 of experts architecture. What does that mean 00:03:12.800 --> 00:03:15.300 in practical terms? It's like a team of small 00:03:15.300 --> 00:03:18.360 experts instead of one brain. Instead of activating 00:03:18.360 --> 00:03:21.520 the entire massive brain, it selectively fires. 00:03:21.819 --> 00:03:24.599 It uses a math expert for calculations and a 00:03:24.599 --> 00:03:27.960 writing expert for linguistics. Exactly. It punches 00:03:27.960 --> 00:03:30.280 way above its weight class for logical reasoning. 00:03:30.800 --> 00:03:34.360 Then we reach the flagship 31b large model. That 00:03:34.360 --> 00:03:37.460 requires 32 gigabytes of RAM and a dedicated 00:03:37.460 --> 00:03:41.000 GPU. It is for pro tasks and deep analytical 00:03:41.000 --> 00:03:43.860 reasoning. What exactly happens if I force the 00:03:43.860 --> 00:03:47.639 31B model onto my 8GB laptop? It will completely 00:03:47.639 --> 00:03:50.479 choke, the responses will stagger out painfully 00:03:50.479 --> 00:03:53.080 slow, or it will just freeze. So picking a model 00:03:53.080 --> 00:03:55.280 that's too big ruins the whole experience. It 00:03:55.280 --> 00:03:56.979 really does, so you should definitely test drive 00:03:56.979 --> 00:03:59.099 it first. Downloading a massive file to test 00:03:59.099 --> 00:04:01.479 its writing style seems inefficient. You want 00:04:01.479 --> 00:04:03.400 to verify that the logic aligns with your workflow. 00:04:03.500 --> 00:04:06.159 Thankfully you can evaluate the 26B model online 00:04:06.159 --> 00:04:08.460 right now. You can access it through Google AI 00:04:08.460 --> 00:04:10.860 Studio in your browser. You bypass the heavy 00:04:10.860 --> 00:04:13.159 local hardware requirements entirely for testing. 00:04:13.319 --> 00:04:15.539 You just navigate to the AI Studio dashboard. 00:04:15.740 --> 00:04:17.819 Ignore the overwhelming write panel with all 00:04:17.819 --> 00:04:20.139 the developer settings. You simply switch the 00:04:20.139 --> 00:04:24.740 model from Gemini to Gemma 4 26b. From there, 00:04:24.899 --> 00:04:27.040 you interact with it naturally in the chat window. 00:04:27.379 --> 00:04:30.100 This is the ideal environment to test its visual 00:04:30.100 --> 00:04:33.279 processing. You can upload a photo of a messy, 00:04:33.560 --> 00:04:36.540 handwritten grocery list. And ask the AI to categorize 00:04:36.540 --> 00:04:40.160 items into dairy, veggies, and snacks. It handles 00:04:40.160 --> 00:04:42.620 overlapping ink and terrible handwriting with 00:04:42.620 --> 00:04:46.139 shocking accuracy. Does testing it online accurately 00:04:46.139 --> 00:04:48.839 reflect how it'll feel locally? Yes. The logic 00:04:48.839 --> 00:04:51.300 is identical. It just saves you the initial massive 00:04:51.300 --> 00:04:53.540 download time. It's a perfect test drive before 00:04:53.540 --> 00:04:55.379 committing your hard drive space. Exactly. So 00:04:55.379 --> 00:04:56.899 how do you actually get it on your computer? 00:04:57.139 --> 00:04:58.839 Most assume you need a computer science degree 00:04:58.839 --> 00:05:01.620 for this. But today, you just need a dedicated 00:05:01.620 --> 00:05:04.300 environment called Olama. Just like you need 00:05:04.300 --> 00:05:07.180 VLC to play a movie, you need Olama to run an 00:05:07.180 --> 00:05:09.480 AI model. It packages the incredibly complex 00:05:09.480 --> 00:05:12.379 backend into a clean, unified installer. The 00:05:12.379 --> 00:05:14.579 installation process genuinely takes under three 00:05:14.579 --> 00:05:17.279 minutes now. If you are on Windows, you download 00:05:17.279 --> 00:05:19.480 the executable file. You click Next and look 00:05:19.480 --> 00:05:21.980 for the Olama icon in your Taskar. On a Mac, 00:05:22.199 --> 00:05:24.459 you unzip it and drag it to Applications. You 00:05:24.459 --> 00:05:27.000 just click Open to trust the app. Linux users 00:05:27.000 --> 00:05:29.800 simply paste a single curl command into the terminal. 00:05:30.199 --> 00:05:32.980 The script autonomously fetches the right binaries 00:05:32.980 --> 00:05:36.180 and configures everything. Once Olama is running, 00:05:36.459 --> 00:05:38.620 you download the model inside the app. You just 00:05:38.620 --> 00:05:42.560 type olama pull gemma4 .e4b into your terminal. 00:05:42.740 --> 00:05:45.040 And you can type olamalist to verify the download 00:05:45.040 --> 00:05:47.779 worked. Is this really as simple as installing 00:05:47.779 --> 00:05:50.420 a web browser? Absolutely. The installer handles 00:05:50.420 --> 00:05:52.500 all the heavy lifting behind the scenes automatically. 00:05:52.660 --> 00:05:54.360 So you really don't need to be a developer to 00:05:54.360 --> 00:05:56.660 install this? Not at all. So the model is installed 00:05:56.660 --> 00:05:58.800 and blinking at you. How do you make it actually 00:05:58.800 --> 00:06:01.540 useful for everyday tasks? You have to provide 00:06:01.540 --> 00:06:05.600 highly specific contextual boundaries. They inputs 00:06:05.600 --> 00:06:08.079 mathematically guarantee generic, hallucinated 00:06:08.079 --> 00:06:11.019 outputs. I still wrestle with prompt drift myself, 00:06:11.199 --> 00:06:13.560 getting lazy with my instructions. It is a very 00:06:13.560 --> 00:06:15.899 easy habit to fall into. You don't just ask the 00:06:15.899 --> 00:06:18.259 model how to start gardening. Right. You say 00:06:18.259 --> 00:06:20.540 you have a small balcony with four hours of sun. 00:06:20.800 --> 00:06:23.379 Give me three easy vegetables and pots and watering 00:06:23.379 --> 00:06:26.620 schedules. By constraining variables, you force 00:06:26.620 --> 00:06:30.139 the algorithm to filter out noise. It works beautifully 00:06:30.139 --> 00:06:32.660 for complex logistical planning as well. I mean, 00:06:32.899 --> 00:06:35.100 planning a multi -stop Monday to save gas is 00:06:35.100 --> 00:06:37.579 a great example. You have a school run at 8 a 00:06:37.579 --> 00:06:40.300 .m., a meeting at 10 a .m., gym, and groceries. 00:06:40.379 --> 00:06:42.279 It gives you the most mathematically efficient 00:06:42.279 --> 00:06:45.339 driving route. It is also a profoundly capable 00:06:45.339 --> 00:06:47.800 tool for independent learning. You can ask it 00:06:47.800 --> 00:06:51.879 for HTML, CSS, and JS for a to -do list in a 00:06:51.879 --> 00:06:56.459 single file. You save it as index .html and notepad, 00:06:56.740 --> 00:06:58.800 and you have a working offline website in two 00:06:58.800 --> 00:07:01.480 minutes. It bypasses the entire nightmare of 00:07:01.480 --> 00:07:04.019 configuring local servers. Why does putting the 00:07:04.019 --> 00:07:06.480 code in one file matter for a beginner? Because 00:07:06.480 --> 00:07:08.500 you don't need to link multiple files together. 00:07:08.660 --> 00:07:10.759 You just double -click and it works. It completely 00:07:10.759 --> 00:07:12.839 removes the friction of learning web development 00:07:12.839 --> 00:07:15.019 setup. It really does. But what happens when 00:07:15.019 --> 00:07:17.600 you push the AI with hard logic? This is where 00:07:17.600 --> 00:07:19.920 we hit the edge of language model architecture. 00:07:20.160 --> 00:07:22.620 They are not actually reasoning. They are calculating 00:07:22.620 --> 00:07:25.019 probabilities. Right. So let's look at the 100 00:07:25.019 --> 00:07:27.720 student bus and van puzzle. You have different 00:07:27.720 --> 00:07:31.160 capacities and costs and a strict no empty seats 00:07:31.160 --> 00:07:34.399 rule. The AI might hyper focus on cheapness and 00:07:34.399 --> 00:07:37.319 forget the no empty seats constraint. It commits 00:07:37.319 --> 00:07:39.439 to an answer before doing the sequential math. 00:07:39.639 --> 00:07:42.920 Exactly. You can fix it by correcting it in plain 00:07:42.920 --> 00:07:45.360 language, but the real trick is humanizing the 00:07:45.360 --> 00:07:48.759 tone first. You can ask for a cookie recipe using 00:07:48.759 --> 00:07:52.220 cold butter. but told in the warm tone of a grandmaster 00:07:52.220 --> 00:07:55.139 chef teaching a beginner. It completely shifts 00:07:55.139 --> 00:07:57.839 the response from a sterile list to an engaging 00:07:57.839 --> 00:08:00.959 lesson. For logic puzzles, you use the magic 00:08:00.959 --> 00:08:03.019 chain of thought prompt. You tell it to think 00:08:03.019 --> 00:08:05.220 step by step before you give me the final answer. 00:08:05.480 --> 00:08:08.220 That phrase completely alters the token generation 00:08:08.220 --> 00:08:11.160 mechanism. It evaluates the van capacities out 00:08:11.160 --> 00:08:13.639 loud, catching the violation. If it messes up 00:08:13.639 --> 00:08:15.459 the math puzzle, do I need to start a whole new 00:08:15.459 --> 00:08:18.279 chat? No, just reply and tell it exactly which 00:08:18.279 --> 00:08:20.879 rule it broke. It course -corrects. Just treat 00:08:20.879 --> 00:08:23.060 it like a human and point out the exact mistake. 00:08:23.319 --> 00:08:25.839 Exactly. Of course, running cutting -edge tech 00:08:25.839 --> 00:08:28.439 locally isn't always flawless. You are running 00:08:28.439 --> 00:08:30.720 server -grade technology on your personal machine. 00:08:31.199 --> 00:08:33.299 Let's cover the quick fixes for the three most 00:08:33.299 --> 00:08:35.759 common roadblocks. If the tech's generation is 00:08:35.759 --> 00:08:38.080 very slow, the model is too big, you should switch 00:08:38.080 --> 00:08:41.519 from the 26B model down to the E4B model. Or 00:08:41.519 --> 00:08:43.399 you might just have too many background apps 00:08:43.399 --> 00:08:46.049 open. The second issue is the dreaded model, 00:08:46.330 --> 00:08:49.250 not found terminal error. You have to check your 00:08:49.250 --> 00:08:52.389 spelling and tags exactly. You must type gemma 00:08:52.389 --> 00:08:56.470 4 .31b exactly as formatted in the repository. 00:08:57.190 --> 00:08:59.809 Finally, users frequently break the multimodal 00:08:59.809 --> 00:09:02.350 image processing feature entirely. Make sure 00:09:02.350 --> 00:09:05.129 you aren't using a specific text -only download 00:09:05.129 --> 00:09:08.029 tag. And you must use an app UI that supports 00:09:08.029 --> 00:09:11.139 image drag and drop. Yeah, not just a basic text 00:09:11.139 --> 00:09:13.480 terminal. Can I drag an image directly into my 00:09:13.480 --> 00:09:16.179 Mac's terminal window? No, the raw terminal doesn't 00:09:16.179 --> 00:09:18.340 read image files. You need the Elama desktop 00:09:18.340 --> 00:09:20.820 window. Got it. Use the actual visual app interface 00:09:20.820 --> 00:09:23.220 for uploading your images. Exactly. It handles 00:09:23.220 --> 00:09:26.259 the heavy lifting of translating the image. Let's 00:09:26.259 --> 00:09:29.200 synthesize the journey we have taken today. Gemma 00:09:29.200 --> 00:09:31.759 4 represents a massive paradigm shift in computing. 00:09:32.000 --> 00:09:34.799 You don't need a massive server farm to have 00:09:34.799 --> 00:09:37.740 high level AI assistance. You don't need to pay 00:09:37.740 --> 00:09:40.200 $20 a month or have an internet connection. Your 00:09:40.200 --> 00:09:43.440 data stays yours and you dictate the rules. The 00:09:43.440 --> 00:09:46.200 barrier to entry has completely vanished. We 00:09:46.200 --> 00:09:49.000 encourage you to go download the E4B model. Start 00:09:49.000 --> 00:09:51.399 pushing its limits on your own machine today. 00:09:51.539 --> 00:09:54.460 Give it complex constraints and force it to think 00:09:54.460 --> 00:09:57.120 step by step. Push back on its assumptions and 00:09:57.120 --> 00:10:00.809 watch it course correct in real time. Beat. If 00:10:00.809 --> 00:10:02.909 a private offline intelligence is sitting on 00:10:02.909 --> 00:10:05.710 your desk today, how does that change what you're 00:10:05.710 --> 00:10:08.190 capable of building tomorrow entirely off the 00:10:08.190 --> 00:10:08.330 grid?