WEBVTT 00:00:00.000 --> 00:00:02.560 Not long ago, running open source AI felt completely 00:00:02.560 --> 00:00:05.540 out of reach. Now, that barrier has vanished 00:00:05.540 --> 00:00:08.000 entirely. The question isn't how to run it anymore. 00:00:08.140 --> 00:00:10.880 The question is where to run it. So you finally 00:00:10.880 --> 00:00:13.160 stop renting your intelligence. Exactly. You 00:00:13.160 --> 00:00:15.939 want to own that cognitive power. Welcome to 00:00:15.939 --> 00:00:18.239 the Deep Dive. We are really thrilled to have 00:00:18.239 --> 00:00:20.460 you with us today. our mission here is clear 00:00:20.460 --> 00:00:23.420 we are mapping out the definitive march 2026 00:00:23.420 --> 00:00:26.899 protocol specifically for deploying open source 00:00:26.899 --> 00:00:29.820 ai we are looking closely at max ann's latest 00:00:29.820 --> 00:00:32.539 guide it covers running models locally hosted 00:00:32.539 --> 00:00:35.390 and in production The landscape right now is 00:00:35.390 --> 00:00:37.929 honestly just staggering. The defining models 00:00:37.929 --> 00:00:41.130 of early 2026 are incredibly capable. We were 00:00:41.130 --> 00:00:43.350 talking about models like Lama 4 Scout. Yeah, 00:00:43.409 --> 00:00:45.850 the 17 billion parameter version is a perfect 00:00:45.850 --> 00:00:48.549 example. And Mistralarge 3 is another absolute 00:00:48.549 --> 00:00:51.270 heavyweight. Plus you have the entire QN3 .5 00:00:51.270 --> 00:00:53.390 series. These models routinely match proprietary 00:00:53.390 --> 00:00:55.729 giants like GPT -4 now. They just do it at a 00:00:55.729 --> 00:00:57.689 fraction of the cost, which changes the math 00:00:57.689 --> 00:01:00.350 for everyone. Okay, let's unpack this. We are 00:01:00.350 --> 00:01:03.289 going to explore three dominant paths today. 00:01:03.689 --> 00:01:07.510 Local, hosted, and production setups. The goal 00:01:07.510 --> 00:01:11.049 is to give you two massive advantages. Data sovereignty. 00:01:11.849 --> 00:01:14.069 and architectural flexibility. That flexibility 00:01:14.069 --> 00:01:17.370 is the ultimate superpower for builders. You 00:01:17.370 --> 00:01:19.709 completely avoid being locked into a single vendor. 00:01:19.849 --> 00:01:22.609 You dodge their sudden pricing changes or API 00:01:22.609 --> 00:01:25.250 outages. Let's start at the most private level 00:01:25.250 --> 00:01:27.230 possible. Running these models right on your 00:01:27.230 --> 00:01:29.829 own desk, we need to look here before even touching 00:01:29.829 --> 00:01:32.030 the cloud. Right. This is option one in the guide, 00:01:32.530 --> 00:01:35.109 Alama. It is the default starting point for a 00:01:35.109 --> 00:01:37.510 very good reason. Think of Alama like a private 00:01:37.510 --> 00:01:39.609 version of Netflix, but here you actually own 00:01:39.609 --> 00:01:42.000 the hard drive. And you own the TV. The appeal 00:01:42.000 --> 00:01:43.959 of that is incredibly obvious. First, you get 00:01:43.959 --> 00:01:46.280 absolute privacy. Your proprietary data never 00:01:46.280 --> 00:01:48.560 leaves your physical machine. Exactly. Then you 00:01:48.560 --> 00:01:51.739 get highly predictable speed and zero per token 00:01:51.739 --> 00:01:54.599 fees for repeated heavy use. Plus, as of March 00:01:54.599 --> 00:01:58.200 2026, the tech took a massive leap. Alama natively 00:01:58.200 --> 00:01:59.939 supports agentic loops right out of the box. 00:02:00.140 --> 00:02:01.819 I want to break down how that actually looks. 00:02:02.260 --> 00:02:04.760 Let's say I need a Python script. Instead of 00:02:04.760 --> 00:02:07.379 just asking it to write code, I can ask it to 00:02:07.379 --> 00:02:10.020 analyze a messy spreadsheet. It writes the script 00:02:10.020 --> 00:02:12.439 internally. It runs that script. And if it hits 00:02:12.439 --> 00:02:15.639 an error, it just fixes it. Right. It realizes 00:02:15.639 --> 00:02:18.180 it made a math error. It rewrites its own code. 00:02:18.219 --> 00:02:20.340 And it just hands me a clean chart all while 00:02:20.340 --> 00:02:23.000 I am sitting offline on a train. It makes vibe 00:02:23.000 --> 00:02:26.780 coding radically more accessible. You are basically 00:02:26.780 --> 00:02:28.860 having a conversation with your operating system. 00:02:28.939 --> 00:02:32.580 The model uses your local tools to solve complex 00:02:32.580 --> 00:02:34.979 problems. I mean, it sounds like magic. So I 00:02:34.979 --> 00:02:37.219 have to admit something here. Beat, I still wrestle 00:02:37.219 --> 00:02:40.159 with prompt drift myself, and the idea of managing 00:02:40.159 --> 00:02:43.479 my own hardware feels daunting. Ah, yeah. Prompt 00:02:43.479 --> 00:02:45.780 drift is a notoriously frustrating reality for 00:02:45.780 --> 00:02:48.180 local setups. You tweak an open source model 00:02:48.180 --> 00:02:51.099 slightly to fit your workflow. Suddenly, it completely 00:02:51.099 --> 00:02:53.659 loses the specific tone you loved yesterday. 00:02:53.919 --> 00:02:55.780 It is incredibly annoying. The neural weights 00:02:55.780 --> 00:02:58.400 just shift in unpredictable ways. And your hesitation 00:02:58.400 --> 00:03:01.550 about hardware is entirely justified, too. It 00:03:01.550 --> 00:03:04.050 brings up the major friction point nobody likes 00:03:04.050 --> 00:03:07.710 talking about. Hardware debt. Local models are 00:03:07.710 --> 00:03:10.349 not magically free. Right. Your physical computer 00:03:10.349 --> 00:03:12.849 actually pays the toll. Exactly. If you want 00:03:12.849 --> 00:03:15.650 to run a true frontier model locally, something 00:03:15.650 --> 00:03:18.310 like Lama 4 Maverick with 400 billion parameters, 00:03:18.710 --> 00:03:21.830 you need serious heavy iron on your desk. Wait, 00:03:21.870 --> 00:03:23.629 what does heavy iron actually mean in this context? 00:03:23.710 --> 00:03:26.979 Am I buying a new laptop or a surfer rack? You 00:03:26.979 --> 00:03:29.500 need massive amounts of unified memory. We are 00:03:29.500 --> 00:03:32.379 talking high -end Apple silicon like an M4 or 00:03:32.379 --> 00:03:35.439 M5 Max or custom rig with multiple dedicated 00:03:35.439 --> 00:03:38.699 NVIDIA GPUs. That sounds expensive. It is. If 00:03:38.699 --> 00:03:40.740 you try to run Maverick on a standard machine. 00:03:41.469 --> 00:03:43.870 The latency is physically painful. It might print 00:03:43.870 --> 00:03:46.030 one word every 10 seconds. So if I'm running 00:03:46.030 --> 00:03:48.469 this on a standard laptop today, is the free 00:03:48.469 --> 00:03:51.610 aspect actually a trap? For massive models, yes. 00:03:51.830 --> 00:03:53.830 Standard consumer machines usually max out around 00:03:53.830 --> 00:03:56.569 7 billion parameters comfortably. Anything larger 00:03:56.569 --> 00:03:58.830 becomes unworkably slow. You end up bleeding 00:03:58.830 --> 00:04:00.889 productivity. You pay in lost time instead of 00:04:00.889 --> 00:04:03.689 money. Right. You skip API fees, but pay heavily 00:04:03.689 --> 00:04:06.110 in hardware and waiting time. Which is exactly 00:04:06.110 --> 00:04:08.289 why most builders eventually hit a wall locally. 00:04:08.509 --> 00:04:10.590 Your ambition just outgrows your motherboard. 00:04:10.909 --> 00:04:13.050 And since hardware limits these local setups 00:04:13.050 --> 00:04:16.069 so quickly, we naturally move to the next logical 00:04:16.069 --> 00:04:19.129 solution. One that bypasses your computer's thermal 00:04:19.129 --> 00:04:22.610 limits entirely. Hosted APIs. This is option 00:04:22.610 --> 00:04:25.110 two in the framework. Hugging face inference 00:04:25.110 --> 00:04:27.819 providers. I always picture this like stacking 00:04:27.819 --> 00:04:31.139 Lego blocks of data. It gives you access to thousands 00:04:31.139 --> 00:04:33.899 of models, but you do it without touching a single 00:04:33.899 --> 00:04:35.860 server yourself. What's fascinating here is how 00:04:35.860 --> 00:04:38.180 practical it is. It is the absolute smartest 00:04:38.180 --> 00:04:41.660 middle ground for an MVP. You can test DeepSeq 00:04:41.660 --> 00:04:44.560 v3 .2 for complex mathematical reasoning. Or 00:04:44.560 --> 00:04:47.240 you can test Mistral Small 4 for pure speed. 00:04:47.579 --> 00:04:50.060 And you do all of this using OpenAI compatible 00:04:50.060 --> 00:04:52.649 endpoints. Let's pause for a quick definition 00:04:52.649 --> 00:04:55.509 here. An API endpoint is simply a digital doorway 00:04:55.509 --> 00:04:57.569 where your app sends questions and gets answers. 00:04:57.810 --> 00:05:00.389 Because it uses that standardized doorway, it 00:05:00.389 --> 00:05:02.810 is seamless. You do not rewrite your application 00:05:02.810 --> 00:05:05.569 code. The code you originally wrote for GPT -4 00:05:05.569 --> 00:05:07.750 still works perfectly. You just change the URL 00:05:07.750 --> 00:05:10.769 string and your API key. That is incredibly smooth. 00:05:11.310 --> 00:05:13.370 But there is a very honest trade -off here we 00:05:13.370 --> 00:05:15.649 must acknowledge. Ultimate convenience always 00:05:15.649 --> 00:05:19.209 costs money. You pay per request. As your user 00:05:19.209 --> 00:05:22.050 base grows, your monthly bill grows right alongside 00:05:22.050 --> 00:05:24.769 it. You also surrender a significant amount of 00:05:24.769 --> 00:05:27.589 control. You lose strict guarantees over network 00:05:27.589 --> 00:05:30.550 latency. And you completely give up strict data 00:05:30.550 --> 00:05:32.910 sovereignty. The data is leaving your building. 00:05:33.310 --> 00:05:36.610 At what exact point does paying per request become 00:05:36.610 --> 00:05:39.250 a bad business decision? It becomes a massive 00:05:39.250 --> 00:05:41.370 liability during high -volume production workloads, 00:05:41.610 --> 00:05:43.449 when you have thousands of users hitting the 00:05:43.449 --> 00:05:46.029 app constantly. That per -token pricing scales 00:05:46.029 --> 00:05:48.490 out of control quickly. It is also a non -starter 00:05:48.490 --> 00:05:51.009 for highly regulated environments. Got it. Great 00:05:51.009 --> 00:05:53.029 for fast prototyping, bad for high -volume restrict 00:05:53.029 --> 00:05:56.009 privacy. Mid -roll sponsor break. Welcome back 00:05:56.009 --> 00:05:57.649 to the Deem Talk. We just covered the mechanics 00:05:57.649 --> 00:06:00.509 of hosted APIs. But if those token costs are 00:06:00.509 --> 00:06:03.050 suddenly adding up, And you need total architectural 00:06:03.050 --> 00:06:05.610 control. You basically have to become the provider 00:06:05.610 --> 00:06:08.689 yourself. That brings us to option three, production 00:06:08.689 --> 00:06:12.129 scale. This is where we introduce the heavy artillery, 00:06:12.470 --> 00:06:16.509 VLLM. It is a highly optimized library designed 00:06:16.509 --> 00:06:19.029 specifically for serving massive language models 00:06:19.029 --> 00:06:21.310 in production environments. It handles the really 00:06:21.310 --> 00:06:23.649 brutal math required to serve thousands of users. 00:06:23.949 --> 00:06:26.589 Right. And the secret sauce is how it manages 00:06:26.589 --> 00:06:28.709 continuous batching. I want to make sure we really 00:06:28.709 --> 00:06:31.240 understand how that works. Contrast continuous 00:06:31.240 --> 00:06:33.920 batching with how older systems used to do it. 00:06:34.000 --> 00:06:37.079 Older systems use naive batching. Imagine a short 00:06:37.079 --> 00:06:40.079 order cook. They take five orders. They wait 00:06:40.079 --> 00:06:42.500 until all five meals are completely cooked before 00:06:42.500 --> 00:06:45.040 serving anyone. So if one order is a massive 00:06:45.040 --> 00:06:47.980 steak, the guy who ordered toast waits 20 minutes. 00:06:48.220 --> 00:06:50.759 Exactly. It was terribly inefficient. Continuous 00:06:50.759 --> 00:06:53.910 batching changes the game entirely. The AI processes 00:06:53.910 --> 00:06:56.990 requests token by token. The millisecond a slot 00:06:56.990 --> 00:06:59.529 opens up in the GPU's memory, it slips a new 00:06:59.529 --> 00:07:01.689 request right into the processing pipeline. Nobody 00:07:01.689 --> 00:07:04.329 waits for the stick to finish. Exactly. The guide 00:07:04.329 --> 00:07:06.569 also mentions that VLLM handles quantization 00:07:06.569 --> 00:07:09.250 seamlessly. Let's clarify that term for a moment. 00:07:09.449 --> 00:07:12.470 Quantization means shrinking an AI model's file 00:07:12.470 --> 00:07:15.730 size to use less computer memory. It drops the 00:07:15.730 --> 00:07:17.970 precision of the numbers inside the neural network. 00:07:18.250 --> 00:07:20.889 A massive model becomes much more manageable. 00:07:21.439 --> 00:07:23.819 But it still requires serious computing power 00:07:23.819 --> 00:07:27.139 to run at scale. Whoa. Imagine scaling to a billion 00:07:27.139 --> 00:07:31.139 queries on your own stack. Two sec silence. It 00:07:31.139 --> 00:07:33.699 sounds incredibly empowering. But here is the 00:07:33.699 --> 00:07:37.339 harsh reality check about self -hosting. You 00:07:37.339 --> 00:07:40.819 own everything. You own the cyber security, load 00:07:40.819 --> 00:07:43.360 balancing, the midnight server crashes. It is 00:07:43.360 --> 00:07:45.839 a massive responsibility. Building your own infrastructure 00:07:45.839 --> 00:07:48.579 is like building a power plant. It is way harder 00:07:48.579 --> 00:07:50.420 than just plugging your lamp into the wall. It 00:07:50.420 --> 00:07:52.699 is learnable. But it requires serious technical 00:07:52.699 --> 00:07:54.759 comfort. Here's where it gets really interesting, 00:07:54.879 --> 00:07:57.300 though. The source guide outlines an incredible 00:07:57.300 --> 00:08:00.279 hybrid trick. Hybrid inference. This is what 00:08:00.279 --> 00:08:02.379 highly resourceful, cost -conscious developers 00:08:02.379 --> 00:08:04.540 are doing right now. It is genuinely brilliant. 00:08:04.680 --> 00:08:07.939 They rent a very cheap cloud VPS, a virtual private 00:08:07.939 --> 00:08:10.540 server, through providers like Hetzner or Hostinger. 00:08:10.620 --> 00:08:13.459 They might pay just $5 to $10 a month. The VPS 00:08:13.459 --> 00:08:16.730 acts as the digital storefront. It handles the 00:08:16.730 --> 00:08:19.589 web traffic and the SSL certificates. But they 00:08:19.589 --> 00:08:22.810 do not process the heavy AI math on that cheap 00:08:22.810 --> 00:08:25.490 server. No. They route the heavy processing securely 00:08:25.490 --> 00:08:28.990 back home. Yes. They use a secure encrypted tunnel, 00:08:29.129 --> 00:08:32.230 specifically TailScale. They route the API calls 00:08:32.230 --> 00:08:35.490 directly to a powerful local machine, something 00:08:35.490 --> 00:08:38.830 like a Mac Mini M4 sitting under a desk at home. 00:08:39.009 --> 00:08:42.159 It is the ultimate architectural hack. It completely 00:08:42.159 --> 00:08:45.759 bypasses the need to rent expensive cloud GPUs. 00:08:45.779 --> 00:08:47.899 You get the public uptime of a cloud server, 00:08:48.059 --> 00:08:50.480 but the actual cognitive processing power comes 00:08:50.480 --> 00:08:52.860 from hardware you already own. Is building a 00:08:52.860 --> 00:08:55.039 hybrid VPS tunnel something a beginner should 00:08:55.039 --> 00:08:57.360 even attempt? Beginners should definitely avoid 00:08:57.360 --> 00:08:59.620 it. It is strictly for advanced builders, people 00:08:59.620 --> 00:09:01.720 who want cloud convenience without paying hourly 00:09:01.720 --> 00:09:04.580 cloud GPU prices. You need to understand networking 00:09:04.580 --> 00:09:06.899 and firewall rules to pull it off securely. Makes 00:09:06.899 --> 00:09:09.200 sense. Start simple, build the hybrid tunnel 00:09:09.200 --> 00:09:11.759 only when token costs hurt. Exactly. You only 00:09:11.759 --> 00:09:13.879 take on that level of networking complexity when 00:09:13.879 --> 00:09:16.379 it solves a specific financial pain point. We 00:09:16.379 --> 00:09:18.740 have covered the main three paths now, local, 00:09:18.879 --> 00:09:22.139 hosted, and production setups. But to truly understand 00:09:22.139 --> 00:09:24.580 the landscape, we have to look at the extremes. 00:09:25.179 --> 00:09:28.259 Absolute zero setup on one end and extreme local 00:09:28.259 --> 00:09:30.840 optimization on the other. Let's examine absolute 00:09:30.840 --> 00:09:33.960 zero setup first. Browser playgrounds. This is 00:09:33.960 --> 00:09:36.559 the easiest difficulty level available. Think 00:09:36.559 --> 00:09:40.639 of platforms like arena .ai, grok .com, or hugging 00:09:40.639 --> 00:09:43.080 face spaces. You literally just open a browser 00:09:43.080 --> 00:09:44.840 tab and start typing. The friction is entirely 00:09:44.840 --> 00:09:48.240 gone. Google Colab even offers a free T4 GPU 00:09:48.240 --> 00:09:51.639 tier. It is an unbelievable resource for educators. 00:09:51.980 --> 00:09:54.179 Students can run complex Python notebooks at 00:09:54.179 --> 00:09:56.539 no cost. But there is a massive glaring catch 00:09:56.539 --> 00:09:59.159 to all of this. Zero privacy. Absolute zero. 00:09:59.379 --> 00:10:02.120 Your proprietary data goes directly to whoever 00:10:02.120 --> 00:10:04.620 hosts that playground. They use your inputs to 00:10:04.620 --> 00:10:06.899 train their future models. And your custom environments 00:10:06.899 --> 00:10:09.000 just vanish the second your session expires. 00:10:09.120 --> 00:10:11.259 Exactly. Let's contrast that zero friction approach 00:10:11.259 --> 00:10:14.000 with the absolute opposite extreme. Edge AI. 00:10:14.179 --> 00:10:16.460 This is the very hard difficulty level. It is 00:10:16.460 --> 00:10:18.299 the bleeding edge of the industry right now. 00:10:18.460 --> 00:10:21.240 This involves packaging the AI models directly 00:10:21.240 --> 00:10:24.889 inside a mobile or desktop app. We see this with 00:10:24.889 --> 00:10:27.710 Apple Intelligence or Gemini Nano. The model 00:10:27.710 --> 00:10:30.350 lives entirely on the silicon inside your phone. 00:10:30.759 --> 00:10:33.139 The theoretical benefits are incredible. You 00:10:33.139 --> 00:10:36.200 get instantaneous responses, full data privacy 00:10:36.200 --> 00:10:38.620 because nothing transmits over the network. And 00:10:38.620 --> 00:10:41.500 absolutely zero network latency. But the technical 00:10:41.500 --> 00:10:44.100 hurdle to actually achieve that is massive. It 00:10:44.100 --> 00:10:46.679 is an engineering nightmare. Compressing these 00:10:46.679 --> 00:10:49.620 highly capable models to run on consumer phones 00:10:49.620 --> 00:10:52.399 is difficult. You must do it without destroying 00:10:52.399 --> 00:10:55.059 the device's battery life. That is the real friction 00:10:55.059 --> 00:10:57.399 point, isn't it? Thermal throttling. Exactly 00:10:57.399 --> 00:10:59.820 the problem. If a phone runs a heavy model constantly, 00:11:00.090 --> 00:11:02.309 The processor heats up, the operating system 00:11:02.309 --> 00:11:04.509 forcefully throttles the chip to prevent melting, 00:11:04.629 --> 00:11:08.490 and your battery drains from 100 to 0 in 20 minutes. 00:11:08.710 --> 00:11:10.590 And because of that compression, the cognitive 00:11:10.590 --> 00:11:13.129 capabilities are usually lower than massive cloud 00:11:13.129 --> 00:11:16.090 models. There is one final advanced path we should 00:11:16.090 --> 00:11:19.149 briefly mention, managed cloud solutions. This 00:11:19.149 --> 00:11:22.059 is strictly enterprise territory. Systems that 00:11:22.059 --> 00:11:24.519 automatically scale server infrastructure during 00:11:24.519 --> 00:11:27.620 massive traffic spikes. Most early -stage projects 00:11:27.620 --> 00:11:30.519 simply do not need this level of expensive complexity. 00:11:30.940 --> 00:11:33.360 With edge AI preserving battery and privacy, 00:11:33.720 --> 00:11:36.200 will it eventually kill cloud APIs entirely? 00:11:36.759 --> 00:11:39.960 Probably not entirely anytime soon. Running massive, 00:11:40.159 --> 00:11:43.200 highly capable frontier models will always require 00:11:43.200 --> 00:11:46.929 significantly more compute power. far more than 00:11:46.929 --> 00:11:49.269 a slim piece of pocket glass can physically hold. 00:11:49.429 --> 00:11:52.389 Edge is for privacy and speed. The cloud remains 00:11:52.389 --> 00:11:55.049 for heavy, complex lifting. They will coexist 00:11:55.049 --> 00:11:57.889 in a hybrid ecosystem, serving very different 00:11:57.889 --> 00:12:00.610 specific cognitive needs. So what does this all 00:12:00.610 --> 00:12:02.970 mean for you? It ultimately comes down to a very 00:12:02.970 --> 00:12:04.970 simple decision tree. You must start with the 00:12:04.970 --> 00:12:06.870 problem you have. Do not start with the technology 00:12:06.870 --> 00:12:09.649 stack. If you desperately need strict data privacy, 00:12:09.990 --> 00:12:12.769 use Alama locally. And if you need rapid iteration 00:12:12.769 --> 00:12:15.950 to validate an idea, you use hugging face inference 00:12:15.950 --> 00:12:18.269 providers. Pay the small token fee for speed. 00:12:18.509 --> 00:12:20.809 And if you need ultimate scale and financial 00:12:20.809 --> 00:12:23.590 control, then you look at VLM and dedicated servers. 00:12:23.809 --> 00:12:26.169 You upgrade your setup only when the need is 00:12:26.169 --> 00:12:29.009 real. That flexibility is the ultimate advantage 00:12:29.009 --> 00:12:31.690 of open source. That flexibility really does 00:12:31.690 --> 00:12:35.159 change everything. Beat. You know, I want to 00:12:35.159 --> 00:12:37.000 leave you with a final thought to mull over today. 00:12:37.159 --> 00:12:40.759 We spent time unpacking edge AI versus the cloud. 00:12:41.340 --> 00:12:44.320 If consumer hardware continues to evolve at this 00:12:44.320 --> 00:12:47.220 rapid pace and quantization techniques get even 00:12:47.220 --> 00:12:49.460 more aggressive, we might see a very strange 00:12:49.460 --> 00:12:52.320 future relatively soon. A future where the massive 00:12:52.320 --> 00:12:55.259 data centers of 2026 become obsolete for daily 00:12:55.259 --> 00:12:58.159 AI tasks. Imagine a world where all collective 00:12:58.159 --> 00:12:59.960 human knowledge fits comfortably on the phone 00:12:59.960 --> 00:13:02.799 in your pocket. Fully offline. Uncensorable. 00:13:03.549 --> 00:13:05.730 What happens to the trillion dollar cloud empires 00:13:05.730 --> 00:13:07.649 then? The entire power dynamic of the internet 00:13:07.649 --> 00:13:09.590 would effectively flip overnight. It really could. 00:13:09.769 --> 00:13:12.090 But until that day comes, you have powerful tools 00:13:12.090 --> 00:13:13.990 in front of you right now. Pick one specific 00:13:13.990 --> 00:13:16.509 problem today and just start building. Stop renting 00:13:16.509 --> 00:13:19.029 your intelligence and start owning it. Out your 00:13:19.029 --> 00:13:19.529 own music.