WEBVTT 00:00:00.000 --> 00:00:03.299 Here's the core competitive reality in AI right 00:00:03.299 --> 00:00:05.360 now. You might build something brilliant really 00:00:05.360 --> 00:00:08.300 fast, but if it's just wrapping a generic API, 00:00:08.660 --> 00:00:12.519 well, the big players, Google, OpenAI, they can 00:00:12.519 --> 00:00:14.519 clone you basically overnight. Yeah, it's like 00:00:14.519 --> 00:00:16.140 you're building on rented land. It feels like 00:00:16.140 --> 00:00:19.719 you're innovating, but there's no real competitive 00:00:19.719 --> 00:00:22.339 edge there. So today, we're diving into the thing 00:00:22.339 --> 00:00:25.079 that actually breaks that cycle. Building proprietary 00:00:25.079 --> 00:00:28.719 AI tech through fine -tuning. Exactly. Welcome 00:00:28.719 --> 00:00:30.920 to the Deep Dive. Our mission today is really 00:00:30.920 --> 00:00:33.840 for you, the listener. We want to move past just 00:00:33.840 --> 00:00:36.280 being an AI user. We're going to lay out how 00:00:36.280 --> 00:00:38.500 you become an AI trainer. We're looking at a 00:00:38.500 --> 00:00:40.799 guide on building these unique high -performance 00:00:40.799 --> 00:00:43.320 models and doing it fast sometimes, like under 00:00:43.320 --> 00:00:46.380 15 minutes using free tools. We'll unpack the 00:00:46.380 --> 00:00:48.119 strategy behind it, you know, better performance, 00:00:48.359 --> 00:00:50.140 real independence. Yeah, and we'll look at the 00:00:50.140 --> 00:00:52.200 tools, the specific open -source models, the 00:00:52.200 --> 00:00:54.240 data you absolutely need. And then walk through 00:00:54.240 --> 00:00:56.759 the practical steps. The goal here is taking 00:00:56.759 --> 00:00:59.250 a big jump. general model and turning it into 00:00:59.250 --> 00:01:02.689 a world champion specialist for whatever your 00:01:02.689 --> 00:01:04.909 specific need is. Let's kick things off with 00:01:04.909 --> 00:01:09.170 that strategy part. Why fine tuning? Why is it 00:01:09.170 --> 00:01:11.349 the defense you need? Well, like we said, most 00:01:11.349 --> 00:01:13.469 startups today, they're kind of disposable if 00:01:13.469 --> 00:01:16.230 they're just API wrappers. The giants see a successful 00:01:16.230 --> 00:01:18.590 feature, they just add it to their own platform. 00:01:18.969 --> 00:01:21.799 Poof. Start it's gone. Fine -tuning, though, 00:01:21.900 --> 00:01:24.299 that's different. It's moving from using a commodity 00:01:24.299 --> 00:01:27.000 to owning something proprietary. Totally. Custom 00:01:27.000 --> 00:01:29.819 models are built different. Unique data, trained 00:01:29.819 --> 00:01:32.799 for specific things. That builds a real defensible 00:01:32.799 --> 00:01:35.239 moat. Something proprietary that a giant can't 00:01:35.239 --> 00:01:37.340 just copy easily. Exactly. They can't just flip 00:01:37.340 --> 00:01:39.840 a switch in their next update and replicate your 00:01:39.840 --> 00:01:42.290 specific model. And you see investors noticing 00:01:42.290 --> 00:01:45.310 this, too. The sources point out that big accelerators 00:01:45.310 --> 00:01:47.989 like Y Combinator, they're looking for founders 00:01:47.989 --> 00:01:50.769 building exactly these kinds of businesses. They 00:01:50.769 --> 00:01:53.390 see the potential there. Monopoly profits potentially. 00:01:53.989 --> 00:01:56.549 That's a huge signal from the market. OK, so 00:01:56.549 --> 00:02:00.390 let's define it simply. Fine tuning is what exactly? 00:02:00.670 --> 00:02:03.189 It's basically adjusting a pre -trained large 00:02:03.189 --> 00:02:05.310 language model. Yeah. One of those big general 00:02:05.310 --> 00:02:08.250 AIs to tweak its internal knobs, its weights. 00:02:08.800 --> 00:02:11.259 And you do that to make it better at very specific, 00:02:11.379 --> 00:02:14.500 narrow tasks. Right. Improve performance just 00:02:14.500 --> 00:02:17.400 where you need it. And that leads to this pretty 00:02:17.400 --> 00:02:20.599 amazing performance claim that a small, fine 00:02:20.599 --> 00:02:24.120 -tuned model can sometimes beat the huge, general 00:02:24.120 --> 00:02:27.740 ones, like even a future GPT -5. On those specialized 00:02:27.740 --> 00:02:30.219 tasks, yeah. It's like taking a really talented 00:02:30.219 --> 00:02:32.879 athlete and training them to be, like, the world's 00:02:32.879 --> 00:02:34.800 best swimmer. Instead of just generally good 00:02:34.800 --> 00:02:38.340 at sports. Wow. Yeah. Imagine that. A 20 billion 00:02:38.340 --> 00:02:40.860 parameter model beating a trillion parameter 00:02:40.860 --> 00:02:44.099 giant on your specific thing. That's the power. 00:02:44.819 --> 00:02:47.139 Specialization gives you this huge return. So 00:02:47.139 --> 00:02:48.919 give me an example. Like what kind of specialized 00:02:48.919 --> 00:02:51.199 task are we talking about? Analyzing specific 00:02:51.199 --> 00:02:54.360 medical images faster maybe? Or understanding 00:02:54.360 --> 00:02:56.879 really niche legal jargon? Precisely that kind 00:02:56.879 --> 00:02:59.219 of thing. Fine tuning lets the model really get 00:02:59.219 --> 00:03:01.960 the nuances, the subtext, the jargon in, say. 00:03:02.400 --> 00:03:04.919 insurance claims processing or maybe some obscure 00:03:04.919 --> 00:03:07.039 programming language. Things where general models 00:03:07.039 --> 00:03:09.060 might hallucinate or just get it wrong. Exactly. 00:03:09.219 --> 00:03:11.219 It cuts down errors dramatically where accuracy 00:03:11.219 --> 00:03:13.900 is absolutely critical. The sources also mentioned 00:03:13.900 --> 00:03:18.439 this idea of strategic control, the uncensored 00:03:18.439 --> 00:03:20.139 revolution. Yeah, that's about independence. 00:03:21.000 --> 00:03:22.800 Fine tuning lets you control the content rules, 00:03:23.020 --> 00:03:25.460 the biases. You can build models that align with 00:03:25.460 --> 00:03:27.300 your specific values, not some big corporations. 00:03:27.680 --> 00:03:30.419 You're not stuck with one dominant AI worldview 00:03:30.419 --> 00:03:33.259 dictating everything. Right. It puts the power, 00:03:33.300 --> 00:03:35.240 the control back in the hands of the builder. 00:03:35.689 --> 00:03:38.530 But, OK, doesn't that open the door to, you know, 00:03:38.569 --> 00:03:41.650 models fine tuned for bad stuff? If the goal 00:03:41.650 --> 00:03:44.310 is zero restrictions, how does that balance out? 00:03:44.509 --> 00:03:46.930 Well, the sources really emphasize the need for 00:03:46.930 --> 00:03:49.110 that independence, noting that control itself 00:03:49.110 --> 00:03:52.270 is power. The responsibility for alignment, for 00:03:52.270 --> 00:03:54.689 making sure it's used ethically, that shifts 00:03:54.689 --> 00:03:57.009 entirely to whoever creates the model. Right. 00:03:57.439 --> 00:03:59.800 It moves the guardrails away from one central 00:03:59.800 --> 00:04:03.139 place. So if the benefits are so clear, the performance, 00:04:03.460 --> 00:04:06.240 the moat, the independence, what's the biggest 00:04:06.240 --> 00:04:09.000 thing stopping a regular AI user from becoming 00:04:09.000 --> 00:04:11.819 an AI trainer? What's the main hurdle? It's getting 00:04:11.819 --> 00:04:14.159 beyond just writing prompts. It's actually shaping 00:04:14.159 --> 00:04:17.259 the AI's core knowledge itself. And this skill, 00:04:17.420 --> 00:04:20.540 becoming an AI trainer, that's becoming really 00:04:20.540 --> 00:04:22.259 valuable, right? Absolutely. Most people just 00:04:22.259 --> 00:04:24.459 talk to AI. Fine tuning means you're shaving 00:04:24.459 --> 00:04:26.759 how it fundamentally works. That's a premium 00:04:26.759 --> 00:04:28.879 skill right now. So where do you start? What's 00:04:28.879 --> 00:04:31.480 the base model? Okay. The sources highlight two 00:04:31.480 --> 00:04:34.100 great open source options specifically designed 00:04:34.100 --> 00:04:36.740 for this kind of customization. First, there's 00:04:36.740 --> 00:04:41.439 GPT -OSS 12B. Okay. Smaller, faster, runs surprisingly 00:04:41.439 --> 00:04:43.800 well, maybe even on a good laptop. Then there's 00:04:43.800 --> 00:04:47.060 GPT -OSS 20B. Bigger, more powerful. Yeah, better 00:04:47.060 --> 00:04:49.259 performance potential, but needs more horsepower. 00:04:49.560 --> 00:04:53.259 Think of Mac Studio or cloud GPUs. The key is 00:04:53.259 --> 00:04:54.819 they're meant to be adapted. Hardware's getting 00:04:54.819 --> 00:04:57.300 more accessible, but... The sources say the biggest 00:04:57.300 --> 00:05:00.019 hurdle still is the data. Oh, absolutely. The 00:05:00.019 --> 00:05:02.339 data set. If you want specialized results, you 00:05:02.339 --> 00:05:04.199 need specialized data. Garbage in, garbage out 00:05:04.199 --> 00:05:07.079 is like 10 times truer here. Look at the agent 00:05:07.079 --> 00:05:09.579 felon data set. That's a perfect example of really 00:05:09.579 --> 00:05:12.339 high quality specialized data. It teaches what's 00:05:12.339 --> 00:05:15.040 called agentic behavior. Agentic, meaning it 00:05:15.040 --> 00:05:19.939 can act like reason, plan, use tools. Exactly. 00:05:19.959 --> 00:05:22.540 Like calling an external API to get information 00:05:22.540 --> 00:05:25.629 or perform an action. So this is how you build 00:05:25.629 --> 00:05:27.870 those AI assistants that feel more autonomous, 00:05:27.970 --> 00:05:30.329 like what people think the big companies use 00:05:30.329 --> 00:05:33.990 for, say, GPT -5's agent mode. Very likely something 00:05:33.990 --> 00:05:36.430 similar, yeah. And the structure of that data 00:05:36.430 --> 00:05:38.990 is critical. How so? Well, these high -quality 00:05:38.990 --> 00:05:41.449 data sets, they follow a specific conversational 00:05:41.449 --> 00:05:44.810 pattern, usually alternating between a user prompt 00:05:44.810 --> 00:05:47.920 and an assistant response. often in a format 00:05:47.920 --> 00:05:51.300 called JSON. Ah, okay. So that structure itself 00:05:51.300 --> 00:05:53.379 teaches the model the right way to interact. 00:05:53.699 --> 00:05:55.579 You got it. It learns the pattern, the style 00:05:55.579 --> 00:05:58.220 you want. So with Agent Flan, I could build something 00:05:58.220 --> 00:06:00.600 that doesn't just answer my question, but actually, 00:06:00.779 --> 00:06:03.399 I don't know, books a meeting by calling my calendar 00:06:03.399 --> 00:06:07.139 API safely. That's the idea. Real autonomy, but 00:06:07.139 --> 00:06:09.740 rooted in very specific training. I have to admit, 00:06:09.819 --> 00:06:11.879 I still wrestle with the data cleaning part myself 00:06:11.879 --> 00:06:14.439 sometimes. Getting that JSON perfect, avoiding 00:06:14.439 --> 00:06:17.720 tiny format errors, it can eat up so much time. 00:06:17.839 --> 00:06:20.019 Oh, yeah. It's finicky. But, okay, let's say 00:06:20.019 --> 00:06:22.000 our listener, they have this amazing, clean, 00:06:22.079 --> 00:06:24.939 specialized data. How do they actually start 00:06:24.939 --> 00:06:27.740 training that 20B model without needing, like, 00:06:27.800 --> 00:06:30.100 a massive server room? Right. They need two key 00:06:30.100 --> 00:06:33.279 things, a technique called LORRE and an accessible 00:06:33.279 --> 00:06:36.139 platform like Google Colab that gives free access 00:06:36.139 --> 00:06:39.860 to GPUs. Okay, LORRE. Low rank adaptation. Yep. 00:06:39.920 --> 00:06:41.980 And this brings us to the practical steps. Yeah. 00:06:42.079 --> 00:06:44.540 The guide we're looking at aims for that like 00:06:44.540 --> 00:06:47.480 sub 15 minute training run. Which sounds crazy 00:06:47.480 --> 00:06:50.259 fast. It relies on Unsloth, which is a library 00:06:50.259 --> 00:06:52.800 optimized for memory efficiency, and Google Colab, 00:06:52.800 --> 00:06:55.860 specifically using their free Tesla T4 GPUs. 00:06:55.879 --> 00:06:58.399 So Loray is the magic ingredient here. Why is 00:06:58.399 --> 00:07:00.779 it so important? Because. Fine -tuning the entire 00:07:00.779 --> 00:07:03.839 model, all 20 billion parameters? Yeah. That's 00:07:03.839 --> 00:07:06.180 just way too expensive for most people, computationally, 00:07:06.180 --> 00:07:08.800 time -wise. Right. Loray is super clever. It 00:07:08.800 --> 00:07:11.540 freezes the original huge model weights, then 00:07:11.540 --> 00:07:14.040 it adds these small extra layers, adapter layers. 00:07:14.240 --> 00:07:16.680 Yeah. And you only train those tiny adapters. 00:07:16.800 --> 00:07:19.120 Wait, okay, so if the main model is frozen, does 00:07:19.120 --> 00:07:21.600 the GPU still need to hold all 20 billion parameters 00:07:21.600 --> 00:07:24.000 in memory while training just the small adapters? 00:07:24.139 --> 00:07:26.699 Good question. It does need to hold the bass 00:07:26.699 --> 00:07:30.459 model, yes. But Loray, especially combined with 00:07:30.459 --> 00:07:33.899 other tricks like quantization, it drastically 00:07:33.899 --> 00:07:36.339 cuts down the memory needed for the training 00:07:36.339 --> 00:07:38.980 process itself. That's the bottleneck it solves. 00:07:39.500 --> 00:07:42.939 Ah, I see. So the benefit is huge time savings. 00:07:43.319 --> 00:07:46.480 Minutes or hours, not days. Exactly. And it makes 00:07:46.480 --> 00:07:49.139 it doable on much less powerful hardware, like 00:07:49.139 --> 00:07:51.959 that single T4 GPU you get for free on Colab. 00:07:52.060 --> 00:07:53.939 It really democratized the whole thing. Okay, 00:07:54.000 --> 00:07:55.879 so the steps in the guide seem pretty straightforward 00:07:55.879 --> 00:07:58.759 then. Set up your Colab notebook, connect to 00:07:58.759 --> 00:08:01.819 the free T4. Yep. Install the libraries you need, 00:08:01.899 --> 00:08:04.660 like PyTorch, Hugging Face Transformers, Unsloth. 00:08:04.779 --> 00:08:07.699 Apply Lore. Then the critical part. Load your 00:08:07.699 --> 00:08:10.800 own specialized data, like Agent Phalan, replacing 00:08:10.800 --> 00:08:12.759 the default example. And use those chat templates 00:08:12.759 --> 00:08:15.040 you mentioned. Crucial step. That makes sure 00:08:15.040 --> 00:08:17.000 your data's format perfectly matches what the 00:08:17.000 --> 00:08:19.680 model expects. Skimp on that, and your training 00:08:19.680 --> 00:08:21.740 could be worthless. Garbage formatting equals 00:08:21.740 --> 00:08:24.379 garbage results. Got it. Then you just run the 00:08:24.379 --> 00:08:26.620 training loop. Pretty much. You watch the loss 00:08:26.620 --> 00:08:28.459 reduction metric. You want to see that number 00:08:28.459 --> 00:08:30.740 going down over time. Means it's learning. And 00:08:30.740 --> 00:08:33.570 once it's done, you test it. Compare your new 00:08:33.570 --> 00:08:36.049 fine -tuned model against the original base model. 00:08:36.210 --> 00:08:38.570 Yeah, see the difference. You can often run that 00:08:38.570 --> 00:08:41.409 comparison test right there. Or maybe locally 00:08:41.409 --> 00:08:44.269 using a tool like Allama to run the models. And 00:08:44.269 --> 00:08:46.480 saving the result. You just save those small 00:08:46.480 --> 00:08:49.279 Luray adapters. That's the beauty of it. You 00:08:49.279 --> 00:08:51.080 can save those adapters locally, keep everything 00:08:51.080 --> 00:08:53.580 private, or upload them to the Hugging Face Hub. 00:08:53.700 --> 00:08:56.100 Makes it super easy to share or integrate into 00:08:56.100 --> 00:08:59.259 apps. Sponsor Read Placeholder 60 seconds. All 00:08:59.259 --> 00:09:01.700 right, let's talk deployment and the economics. 00:09:02.039 --> 00:09:05.559 Google Collabs tiers seem useful here. That free 00:09:05.559 --> 00:09:08.779 tier with the Tesla T4, perfect for getting started 00:09:08.779 --> 00:09:11.539 experimenting. Absolutely. Learn the ropes, test 00:09:11.539 --> 00:09:13.950 your data. But if you're serious... Moving towards 00:09:13.950 --> 00:09:16.850 actually using this, that paid tier, around $10 00:09:16.850 --> 00:09:19.429 a month, it unlocks much faster GPUs. Like the 00:09:19.429 --> 00:09:22.710 A100s or TPUs. Yeah, A100s can be like three, 00:09:22.789 --> 00:09:25.009 four times faster, TPUs even more sometimes. 00:09:25.250 --> 00:09:27.350 When you're doing lots of runs or working with 00:09:27.350 --> 00:09:29.590 bigger data sets, that time saving is huge. It 00:09:29.590 --> 00:09:31.750 cuts a 12 -hour training run down to maybe three 00:09:31.750 --> 00:09:33.789 or four hours. The return on investment seems 00:09:33.789 --> 00:09:36.570 pretty clear if time is money. Definitely. Development 00:09:36.570 --> 00:09:38.710 speed matters. Okay, but what about the things 00:09:38.710 --> 00:09:42.879 that go wrong? Pitfalls. Running out of GPU memory 00:09:42.879 --> 00:09:45.440 must be common on the free tier. Oh, yeah. Happens 00:09:45.440 --> 00:09:47.580 all the time. But the fixes are usually straightforward. 00:09:48.139 --> 00:09:51.059 Try reducing your batch size process less data 00:09:51.059 --> 00:09:54.220 at once. Or lower the maximum sequence length. 00:09:54.399 --> 00:09:57.799 Yep. Or use 4 -bit quantization. That loads the 00:09:57.799 --> 00:09:59.600 model weights in a really compressed format, 00:09:59.840 --> 00:10:03.340 saves a ton of VRAM. Unsloth makes this super 00:10:03.340 --> 00:10:05.320 easy. And the other big headache you mentioned. 00:10:05.909 --> 00:10:08.669 Data loading problem. Right. You absolutely must 00:10:08.669 --> 00:10:11.450 tell the code exactly where your custom data 00:10:11.450 --> 00:10:14.509 file is. Like data files train my data, my training 00:10:14.509 --> 00:10:17.149 file dot JSONL. If you don't specify that. It'll 00:10:17.149 --> 00:10:19.230 probably assume some default data set or structure 00:10:19.230 --> 00:10:21.730 and you'll waste hours trying to figure out why 00:10:21.730 --> 00:10:23.409 it's not working or why the results are weird. 00:10:23.769 --> 00:10:26.330 Explicit is better. Good tip. And one last point 00:10:26.330 --> 00:10:28.470 on data. If you need specialized data, but it 00:10:28.470 --> 00:10:31.090 just doesn't exist for some really niche application. 00:10:31.470 --> 00:10:34.629 Then you got to make it. Synthetic data generation. 00:10:35.309 --> 00:10:37.690 Use the powerful models we already have, like 00:10:37.690 --> 00:10:42.350 GPT -4, GPT -5 maybe, CLAWD. Task them with generating 00:10:42.350 --> 00:10:45.049 thousands of high -quality examples tailored 00:10:45.049 --> 00:10:47.370 to your need. And curate them carefully, obvious. 00:10:47.450 --> 00:10:50.090 Of course. But generating that unique data set, 00:10:50.230 --> 00:10:52.370 that itself becomes part of your competitive 00:10:52.370 --> 00:10:55.970 moat. But maybe the biggest long -term advantage 00:10:55.970 --> 00:10:58.070 the sources point to is running these models 00:10:58.070 --> 00:11:01.070 locally. Oh, absolutely. Once you fine -tune 00:11:01.070 --> 00:11:03.190 these models, especially the smaller, more efficient 00:11:03.190 --> 00:11:05.909 ones like that 12B or even 20B with quantization, 00:11:06.230 --> 00:11:08.490 they can often run entirely on your own hardware. 00:11:08.629 --> 00:11:11.309 Like a good MacBook Pro or a Mac Studio. Exactly. 00:11:11.330 --> 00:11:13.870 And the benefits there are massive. Perfect privacy, 00:11:14.070 --> 00:11:15.649 right? Yeah. Data never leaves your machine. 00:11:15.870 --> 00:11:19.590 Yep. Zero dependence on cloud providers. No ongoing 00:11:19.590 --> 00:11:22.759 API costs for inference, ever. That sounds crucial 00:11:22.759 --> 00:11:26.250 for certain industries. Healthcare. Legal. Anywhere 00:11:26.250 --> 00:11:29.250 with really sensitive data. Totally. It unlocks 00:11:29.250 --> 00:11:31.309 huge business opportunities, too. Think vertical 00:11:31.309 --> 00:11:34.370 specific AI like the best AI for analyzing only 00:11:34.370 --> 00:11:37.110 construction contracts. Or enterprise tools built 00:11:37.110 --> 00:11:39.549 on a company's internal knowledge base running 00:11:39.549 --> 00:11:42.129 securely inside their network. Or consumer products 00:11:42.129 --> 00:11:44.269 where privacy is the main selling point. It's 00:11:44.269 --> 00:11:46.570 just a fundamentally stronger position than just 00:11:46.570 --> 00:11:49.570 being another API wrapper. So fine tuning really 00:11:49.570 --> 00:11:52.289 is about establishing that proprietary tech moving 00:11:52.289 --> 00:11:55.120 from. What did you call it? Rented land. Yeah, 00:11:55.179 --> 00:11:57.559 from rented land to owned territory. That's where 00:11:57.559 --> 00:11:59.580 the sustainable advantage lies. Okay, let's boil 00:11:59.580 --> 00:12:02.940 this down. For you, the learner listening right 00:12:02.940 --> 00:12:05.399 now, what are the big takeaways from this deep 00:12:05.399 --> 00:12:07.360 dive? I think there are three key things. First, 00:12:07.539 --> 00:12:10.860 don't underestimate specialization. A fine -tuned 00:12:10.860 --> 00:12:14.000 20B model can beat a giant generalist on its 00:12:14.000 --> 00:12:16.620 specific task. Often it will. Second takeaway. 00:12:16.879 --> 00:12:19.830 Laurie, that technique... Low -rank adaptation 00:12:19.830 --> 00:12:22.570 is what made all this accessible. It lets you 00:12:22.570 --> 00:12:24.649 do serious training on hardware you can actually 00:12:24.649 --> 00:12:27.009 get your hands on, maybe even for free. Right, 00:12:27.070 --> 00:12:29.629 it democratized it. And the third? Data, data, 00:12:29.730 --> 00:12:32.230 data. The quality and the specificity of your 00:12:32.230 --> 00:12:34.330 training data. That's the single biggest factor 00:12:34.330 --> 00:12:37.110 determining your success. Not necessarily the 00:12:37.110 --> 00:12:39.629 raw size of the base model, but the quality of 00:12:39.629 --> 00:12:42.610 the data you feed it. You absolutely need the 00:12:42.610 --> 00:12:45.450 right data. Mm -hmm. The tools are out there, 00:12:45.509 --> 00:12:48.669 mostly free. The knowledge is accessible. like 00:12:48.669 --> 00:12:50.610 in the guides we discussed now really is the 00:12:50.610 --> 00:12:52.870 time to build this kind of defensible advantage 00:12:52.870 --> 00:12:55.950 to shift from being just a user to becoming an 00:12:55.950 --> 00:12:58.529 ai trainer yeah make the leap so here's a final 00:12:58.529 --> 00:13:00.990 thought to leave you with maybe the future of 00:13:00.990 --> 00:13:03.990 ai isn't one single giant model trying to do 00:13:03.990 --> 00:13:06.769 everything maybe it's more like a swarm of hyper 00:13:06.769 --> 00:13:09.929 specialized experts independently controlled 00:13:09.929 --> 00:13:12.750 fine -tuned models running efficiently maybe 00:13:12.750 --> 00:13:15.389 even on your own local hardware so the question 00:13:15.389 --> 00:13:18.460 for you is What specialized problem out there 00:13:18.460 --> 00:13:20.600 is just waiting for your custom AI solution?