1 00:00:00,137 --> 00:00:03,895 Your own AI. With prompts, no prior experience needed. 2 00:00:03,920 --> 00:00:08,520 In just ten minutes: your chat, your apps, and budget control. 3 00:00:09,400 --> 00:00:11,793 Let's get started. We'll use Claude and 4 00:00:11,793 --> 00:00:14,615 Runpod, a service for renting powerful servers 5 00:00:14,640 --> 00:00:18,015 for AI, billed by the second while they're running. 6 00:00:18,040 --> 00:00:20,255 You'll learn without buying that hardware. 7 00:00:20,280 --> 00:00:22,655 Here, we've summarized the setup wait times. 8 00:00:22,680 --> 00:00:24,575 You don't need to know how to code. 9 00:00:24,600 --> 00:00:27,828 If you're already a senior developer, you can follow along too. 10 00:00:27,853 --> 00:00:32,615 You direct things with AI: you ask, Claude prepares, and you check the result before 11 00:00:32,640 --> 00:00:36,800 moving on. If you have a question, ask it to explain the step and adapt it to your 12 00:00:36,840 --> 00:00:40,735 computer. The prompts and files are in the course resources. 13 00:00:40,760 --> 00:00:45,335 Runpod lets you rent high-end hardware online, with powerful graphics cards 14 00:00:45,360 --> 00:00:47,055 and plenty of memory. 15 00:00:47,080 --> 00:00:49,495 You pay for the time you keep the hardware running. 16 00:00:49,520 --> 00:00:52,615 Your computer sends the task and receives the response. 17 00:00:52,640 --> 00:00:54,548 You don't have to buy that hardware. 18 00:00:54,573 --> 00:00:59,455 Here's the key point: even when you don't ask it anything, the time it's running keeps 19 00:00:59,480 --> 00:01:00,375 being billed. 20 00:01:00,400 --> 00:01:02,975 First, go to Runpod and create your account. 21 00:01:03,000 --> 00:01:05,082 Here, we've summarized registration and payment. 22 00:01:05,107 --> 00:01:09,360 To continue, you must be able to access the console, be in the correct account, and have 23 00:01:09,400 --> 00:01:13,322 enough balance. We'll check the hardware price before creating it. 24 00:01:13,347 --> 00:01:16,895 For this test, you don't need to sign up for a plan lasting several months. 25 00:01:16,920 --> 00:01:21,242 Before starting, you need Claude Code installed and signed in to your account. 26 00:01:21,267 --> 00:01:24,335 If you don't have it yet, follow the guide in the resources. 27 00:01:24,360 --> 00:01:27,455 Now open Claude in the course folder and ask it to: 28 00:01:27,480 --> 00:01:31,402 check my computer and install any missing connection tools. 29 00:01:31,427 --> 00:01:33,095 Claude runs the checks; 30 00:01:33,120 --> 00:01:35,482 you review and authorize what's needed. 31 00:01:35,507 --> 00:01:38,015 Claude is paid for separately from Runpod. 32 00:01:38,040 --> 00:01:42,122 Download the course files to your own folder and open Claude from there. 33 00:01:42,147 --> 00:01:46,815 They include the instructions, the chat, and an example tool for creating the clip. 34 00:01:46,840 --> 00:01:50,522 This lets you give short commands without repeating the entire setup. 35 00:01:50,547 --> 00:01:54,548 Ollama, Python, and FFmpeg will be installed on the server; 36 00:01:54,573 --> 00:01:57,775 your computer only needs the connection tools. 37 00:01:57,800 --> 00:02:02,575 Now ask Claude to connect the official Runpod MCP using the address in 38 00:02:02,600 --> 00:02:07,360 the instructions. MCP is the adapter that lets Claude use an external 39 00:02:07,480 --> 00:02:11,002 tool. Here, it can check available hardware and prepare ours. 40 00:02:11,027 --> 00:02:15,162 Claude will install it; complete the access process that opens in your browser. 41 00:02:15,187 --> 00:02:19,375 Check the account before authorizing, and reopen the conversation if prompted. 42 00:02:19,400 --> 00:02:24,175 The first request is: read the instructions, check my tools, 43 00:02:24,200 --> 00:02:26,560 and check my pods and the current price. 44 00:02:26,920 --> 00:02:31,855 Don't create anything yet. Runpod calls the rented machine a pod, with your 45 00:02:31,880 --> 00:02:35,040 programs inside it. When you see that word, think: 46 00:02:35,240 --> 00:02:37,175 my computer on Runpod. 47 00:02:37,200 --> 00:02:39,015 First, we verify access; 48 00:02:39,040 --> 00:02:40,480 then we spend money. 49 00:02:40,600 --> 00:02:45,135 Now ask it to suggest a single pod and show you the memory and price. 50 00:02:45,160 --> 00:02:48,895 We're looking for a GPU with 24 GB of memory. 51 00:02:48,920 --> 00:02:51,828 The GPU is the card that accelerates the model. 52 00:02:51,853 --> 00:02:54,682 Its memory stores what the model needs to work. 53 00:02:54,707 --> 00:02:58,375 RAM runs programs, and the disk stores files. 54 00:02:58,400 --> 00:03:02,495 Our example pricing will use an RTX 4090. 55 00:03:02,520 --> 00:03:06,735 We'll store the model and files on a network disk, separate from the pod. 56 00:03:06,760 --> 00:03:10,535 The disk and GPU must be in the same data center: 57 00:03:10,560 --> 00:03:12,695 the location where that hardware is housed. 58 00:03:12,720 --> 00:03:16,295 That way, we can return the GPU and keep the setup. 59 00:03:16,320 --> 00:03:18,855 That disk continues to incur charges as long as it exists; 60 00:03:18,880 --> 00:03:20,880 we'll include it in the final calculation. 61 00:03:21,000 --> 00:03:25,775 Review the summary: one GPU, the compatible disk, the rate, and the 62 00:03:25,800 --> 00:03:28,775 pod name. Availability changes: 63 00:03:28,800 --> 00:03:33,575 use the option Claude has verified and accept only a quote that suits you. 64 00:03:33,600 --> 00:03:35,640 Then it can create the machine. 65 00:03:35,720 --> 00:03:37,480 That's when billing begins. 66 00:03:37,560 --> 00:03:41,120 Let the setup finish, even though we've shortened the wait here. 67 00:03:41,240 --> 00:03:45,815 Ollama is the program that manages and runs AI models on the machine where you 68 00:03:45,840 --> 00:03:48,280 install it—in this case, our server. 69 00:03:48,360 --> 00:03:53,440 The model will be Qwen 3.5 9B, from the Chinese company Alibaba. 70 00:03:53,480 --> 00:03:58,255 We chose it because it combines text, images, and tools in a size that fits 71 00:03:58,280 --> 00:04:00,320 comfortably within the proposed GPU's capacity. 72 00:04:00,520 --> 00:04:04,800 It's a practical balance for this video, without paying for a huge model. 73 00:04:04,920 --> 00:04:08,655 Before opening the chat, ask Claude to check four things: 74 00:04:08,680 --> 00:04:12,960 the connection, Ollama, the downloaded model, and the GPU in use. 75 00:04:13,160 --> 00:04:16,895 Seeing the machine marked as running doesn't mean it's ready yet. 76 00:04:16,920 --> 00:04:20,095 Continue once all four checks are complete. 77 00:04:20,120 --> 00:04:21,984 This way, you know what you're paying for 78 00:04:21,984 --> 00:04:24,575 and that the model has the computing power you requested. 79 00:04:24,600 --> 00:04:28,880 Now tell it: open the secure tunnel and show me my chat in the browser. 80 00:04:29,160 --> 00:04:34,240 An SSH tunnel connects your computer to the server without exposing the AI 81 00:04:34,280 --> 00:04:37,895 to everyone on the internet. Claude prepares the access key; 82 00:04:37,920 --> 00:04:40,120 the private part stays with you. 83 00:04:40,200 --> 00:04:42,975 Don't paste secret keys into the conversation. 84 00:04:43,000 --> 00:04:44,775 Open the address it gives you. 85 00:04:44,800 --> 00:04:48,695 The page must be served from your pod, even though you view it in your browser. 86 00:04:48,720 --> 00:04:52,775 Type a short question and check that you receive a real response. 87 00:04:52,800 --> 00:04:54,895 The first load may take longer. 88 00:04:54,920 --> 00:04:59,695 Ask another question: during the session, we can keep the model ready to 89 00:04:59,720 --> 00:05:01,600 avoid loading it every time. 90 00:05:01,673 --> 00:05:05,155 To check image processing, use something easy to verify: 91 00:05:05,180 --> 00:05:09,515 three large shapes. Attach the image and ask how many it can see. 92 00:05:09,540 --> 00:05:11,995 Compare its answer with what you see in front of you. 93 00:05:12,020 --> 00:05:14,515 Continue once you've verified the result. 94 00:05:14,540 --> 00:05:18,075 This test helps verify that the image reaches the model; 95 00:05:18,100 --> 00:05:21,500 it doesn't make all its answers correct. 96 00:05:21,760 --> 00:05:25,635 Now ask for a simple function and five tests to check it. 97 00:05:25,660 --> 00:05:30,035 Tell Claude to send the task to your model and run the code in the pod. 98 00:05:30,060 --> 00:05:34,715 Look at the results. During setup, we already tested this workflow with a small 99 00:05:34,740 --> 00:05:36,939 piece of code. Starting this way helps you 100 00:05:36,939 --> 00:05:39,715 recognize what the AI requested and what the computer 101 00:05:39,740 --> 00:05:40,780 actually did. 102 00:05:40,820 --> 00:05:42,900 We can also give it tools. 103 00:05:42,980 --> 00:05:44,715 The AI requests an action; 104 00:05:44,740 --> 00:05:49,995 a program executes it. During setup, we tested an FFmpeg tool 105 00:05:50,020 --> 00:05:52,220 to produce a five-second clip. 106 00:05:52,500 --> 00:05:57,995 FFmpeg trims videos, converts formats, combines images, and applies filters. 107 00:05:58,020 --> 00:06:02,275 Generating scenes with other video models requires additional setup. 108 00:06:02,300 --> 00:06:04,555 Here, we'll continue with a simple example. 109 00:06:04,580 --> 00:06:08,995 Ask it to check that the file exists, lasts five seconds, and plays. 110 00:06:09,020 --> 00:06:13,115 The tool must return that result to the model so it can confirm what it 111 00:06:13,140 --> 00:06:16,460 created. Request, action, and result: 112 00:06:16,500 --> 00:06:18,395 those are three distinct stages. 113 00:06:18,420 --> 00:06:21,315 Review all three before marking the task as complete. 114 00:06:21,340 --> 00:06:25,435 This workflow will help you build more useful tasks later. 115 00:06:25,460 --> 00:06:27,995 The chat is one way to use your AI. 116 00:06:28,020 --> 00:06:32,555 Now ask Claude: check that another application can connect and show me the 117 00:06:32,580 --> 00:06:36,060 information it needs. This interface is called an API. 118 00:06:36,100 --> 00:06:38,875 Claude will make the request from the console for you. 119 00:06:38,900 --> 00:06:42,995 A compatible application on this computer can use the address and model 120 00:06:43,020 --> 00:06:44,995 as long as the tunnel remains active. 121 00:06:45,020 --> 00:06:47,980 The tunnel's local address belongs to your computer. 122 00:06:48,100 --> 00:06:52,155 It doesn't automatically become a public address you can use from anywhere 123 00:06:52,180 --> 00:06:56,155 in the world. Today, we're setting up a secure connection for your own use. 124 00:06:56,180 --> 00:07:00,060 With that, you can test your AI through the chat and another program. 125 00:07:00,100 --> 00:07:04,795 When you're finished, ask Claude to delete only the pod for this course and keep 126 00:07:04,820 --> 00:07:08,380 the network disk. Then verify the result in Runpod. 127 00:07:08,660 --> 00:07:11,155 Closing the browser doesn't stop billing. 128 00:07:11,180 --> 00:07:15,595 When you return the GPU, the chat should stop responding because no machine is 129 00:07:15,620 --> 00:07:18,475 running the model anymore. The disk remains. 130 00:07:18,500 --> 00:07:22,795 To return, request an available GPU in the same data center as your disk. 131 00:07:22,820 --> 00:07:27,595 Claude must check the rate, connect that same disk, and test the chat 132 00:07:27,620 --> 00:07:30,715 and API again. This isn't turning the same pod back on: 133 00:07:30,740 --> 00:07:33,820 it's setting up another session using your saved files. 134 00:07:33,860 --> 00:07:37,020 Confirm those checks before continuing to work. 135 00:07:37,300 --> 00:07:41,675 For perspective, we found a machine advertised at six thousand six hundred 136 00:07:41,700 --> 00:07:46,635 sixteen dollars: an RTX 4090 with 24 GB 137 00:07:46,660 --> 00:07:50,940 of graphics memory, 64 GB of RAM, and a 2 TB disk. 138 00:07:50,980 --> 00:07:55,515 This is a purchase price reference for the same GPU as our example, but not the same 139 00:07:55,540 --> 00:07:57,802 pod configuration. Here, you start by renting 140 00:07:57,802 --> 00:08:00,315 computing power by the hour instead of making that 141 00:08:00,340 --> 00:08:04,262 purchase. Whether buying or renting is better depends on how much you use it. 142 00:08:04,287 --> 00:08:08,475 You control who can access your instance, can try models, and connect your 143 00:08:08,500 --> 00:08:13,035 applications. While the machine is running, you can keep experimenting without an 144 00:08:13,060 --> 00:08:15,395 app-imposed message limit: 145 00:08:15,420 --> 00:08:18,755 your limits are the available memory and your budget. 146 00:08:18,780 --> 00:08:21,035 Let's calculate it over thirty days. 147 00:08:21,060 --> 00:08:23,995 Three hours a day adds up to 90 hours a month. 148 00:08:24,020 --> 00:08:28,915 As an example, we checked an RTX 4090 at 89 cents 149 00:08:28,940 --> 00:08:33,675 per hour: that would be $80.10 for the GPU. 150 00:08:33,700 --> 00:08:36,795 This isn't a reservation or a guarantee of availability. 151 00:08:36,820 --> 00:08:39,235 Always check your current quote. 152 00:08:39,260 --> 00:08:43,835 A standard 20 GB network disk adds $1.40 per 153 00:08:43,860 --> 00:08:48,875 month. The machine's temporary 20 GB disk adds about 25 cents 154 00:08:48,900 --> 00:08:50,780 for those 90 hours. 155 00:08:50,820 --> 00:08:54,955 Approximate total: $81.75. 156 00:08:54,980 --> 00:08:56,955 This excludes taxes and Claude. 157 00:08:56,980 --> 00:09:00,620 Downloads and waiting while the machine is running also count. 158 00:09:01,033 --> 00:09:03,500 Before we wrap up, remember this mental map. 159 00:09:03,780 --> 00:09:05,702 It all starts on your computer: 160 00:09:05,727 --> 00:09:07,900 you tell Claude what you need. 161 00:09:07,940 --> 00:09:11,860 Through MCP, Claude checks and prepares the pod on Runpod. 162 00:09:11,940 --> 00:09:17,860 Inside that powerful machine are Ollama, Qwen, and tools such as FFmpeg. 163 00:09:17,940 --> 00:09:20,595 Your computer opens a secure connection: 164 00:09:20,620 --> 00:09:24,060 your chat and the API for other applications come through it. 165 00:09:24,100 --> 00:09:26,035 The model runs on the server; 166 00:09:26,060 --> 00:09:28,060 your computer stays in control. 167 00:09:28,100 --> 00:09:30,620 When you're done, you return the GPU. 168 00:09:30,660 --> 00:09:34,820 If you keep the disk, you save the setup for the next session. 169 00:09:34,900 --> 00:09:37,235 Ask, verify, and shut down: 170 00:09:37,260 --> 00:09:38,460 that's the workflow. 171 00:09:38,810 --> 00:09:40,998 If you've finished testing, download anything 172 00:09:40,998 --> 00:09:43,235 you want to keep and ask to delete the pod and 173 00:09:43,260 --> 00:09:45,275 the network disk for this course. 174 00:09:45,300 --> 00:09:48,075 Verify separately that both are gone. 175 00:09:48,100 --> 00:09:50,380 Keeping the disk lets you return; 176 00:09:50,540 --> 00:09:53,715 deleting it ends that charge and erases its files. 177 00:09:53,740 --> 00:09:56,940 Choose based on your needs and verify the result. 178 00:09:57,044 --> 00:09:58,675 You understand the workflow: 179 00:09:58,700 --> 00:10:02,875 what you need, what you're going to set up, and how to check that it works. 180 00:10:02,900 --> 00:10:07,235 With that clear, you can delegate the setup to Claude and put a powerful server to 181 00:10:07,260 --> 00:10:11,595 work. You'll find the prompts and complete resources in this video's description. 182 00:10:11,620 --> 00:10:14,452 You direct things with AI. 183 00:10:14,477 --> 00:10:17,100 Your next idea already has a place to run.