WEBVTT 00:00:00.000 --> 00:00:02.399 Imagine you are reading a long passage, maybe 00:00:02.399 --> 00:00:06.400 something complex, history or science. Your brain 00:00:06.400 --> 00:00:08.800 doesn't just process word one, then word two 00:00:08.800 --> 00:00:11.080 and completely forget how the beginning connects 00:00:11.080 --> 00:00:13.419 to the end. Exactly right. Your mind is, you 00:00:13.419 --> 00:00:15.820 know, constantly mapping relationships. When 00:00:15.820 --> 00:00:17.739 you read something like the black cat set on 00:00:17.739 --> 00:00:20.839 the mat, you instantly know black describes cat. 00:00:20.839 --> 00:00:23.379 Right. That natural human thing deciding which 00:00:23.379 --> 00:00:25.940 parts are important to other parts. That's attention. 00:00:26.160 --> 00:00:30.309 And that that basic idea. It seems so simple, 00:00:30.530 --> 00:00:32.670 but that's the critical insight. The thing that 00:00:32.670 --> 00:00:36.509 launched all the AI we use now, ChatGPT, Gemini. 00:00:36.630 --> 00:00:40.130 It really is. And look, the pace of new AI names, 00:00:40.350 --> 00:00:43.229 new tools, it can feel completely overwhelming 00:00:43.229 --> 00:00:45.070 if you're trying to keep up. Yeah, it's a lot. 00:00:45.469 --> 00:00:47.460 So our goal today is pretty simple. Let's get 00:00:47.460 --> 00:00:49.219 some foundational clarity. Think of this deep 00:00:49.219 --> 00:00:51.700 dive like understanding the scientific blueprints 00:00:51.700 --> 00:00:54.079 for these huge systems. OK, so we've looked at 00:00:54.079 --> 00:00:55.979 the source material, and it really boils down 00:00:55.979 --> 00:00:58.859 to four big shifts that made modern AI possible. 00:00:59.240 --> 00:01:02.740 First, building the core engine. That's attention. 00:01:03.200 --> 00:01:06.379 Then the scale needed for new powers, which is 00:01:06.379 --> 00:01:08.659 few -shot learning. Getting bigger. How we made 00:01:08.659 --> 00:01:12.099 them helpful and safe alignment. Crucial step. 00:01:12.359 --> 00:01:14.500 And then how they actually connect and interact 00:01:14.500 --> 00:01:17.540 with the real world. That's RIG and agents. So 00:01:17.540 --> 00:01:19.620 let's unpack this. All right, let's start at 00:01:19.620 --> 00:01:22.260 the beginning. The cornerstone paper from 2017, 00:01:22.400 --> 00:01:24.180 the one that introduced the transformer architecture. 00:01:24.980 --> 00:01:28.239 Before this, AI had this, well, fundamental memory 00:01:28.239 --> 00:01:31.140 problem. Oh, huge structural problem. Yeah. Older 00:01:31.140 --> 00:01:33.280 AI models, things like recurrent neural networks 00:01:33.280 --> 00:01:37.859 or RNNs, they read text sequentially, like reading 00:01:37.859 --> 00:01:40.459 a scroll, word one, then word two. OK. By the 00:01:40.459 --> 00:01:43.200 time they got to maybe the 50th word or the 100th, 00:01:43.439 --> 00:01:47.019 the computational memory, it just faded. They 00:01:47.019 --> 00:01:48.579 forgot the start of the sentence. Meaning they 00:01:48.579 --> 00:01:51.219 couldn't realistically summarize a long document 00:01:51.219 --> 00:01:54.640 or translate a complex paragraph accurately because 00:01:54.640 --> 00:01:56.599 they lost the context. Exactly, they lost the 00:01:56.599 --> 00:01:58.640 context. The transformer totally changed this. 00:01:58.980 --> 00:02:01.799 It allowed the model to look at all the words 00:02:01.799 --> 00:02:04.159 in a sentence basically simultaneously. All at 00:02:04.159 --> 00:02:07.340 once. All at once. Which allows for massive parallel 00:02:07.340 --> 00:02:10.240 processing. It can instantly connect the beginning 00:02:10.240 --> 00:02:13.629 of a really long sentence with the end. and the 00:02:13.629 --> 00:02:16.229 key mechanism is called self -attention. Okay, 00:02:16.310 --> 00:02:18.069 explain self -attention again. You had a good 00:02:18.069 --> 00:02:19.789 analogy for this. Right. Think about being at 00:02:19.789 --> 00:02:23.789 a noisy party. Okay. You're talking to your friend, 00:02:24.389 --> 00:02:27.490 but there are dozens of other conversations swirling 00:02:27.490 --> 00:02:30.610 around. Your brain uses self -tension to filter 00:02:30.610 --> 00:02:32.969 out all that noise and focus just on your friend's 00:02:32.969 --> 00:02:36.949 voice. The AI does basically the same thing. 00:02:36.969 --> 00:02:39.050 For every single word it processes, it gives 00:02:39.050 --> 00:02:42.039 every other word an important score. It builds 00:02:42.039 --> 00:02:45.340 this like instantaneous map of relationships. 00:02:45.620 --> 00:02:47.379 So instead of just step by step, it's building 00:02:47.379 --> 00:02:49.780 a whole web of connections for everything it 00:02:49.780 --> 00:02:52.580 sees. That sounds incredibly powerful. And it 00:02:52.580 --> 00:02:54.539 is, I mean, it's the foundation for pretty much 00:02:54.539 --> 00:02:56.819 every large language model today. It is. But 00:02:56.819 --> 00:02:58.680 you mentioned there's a big technical bottleneck 00:02:58.680 --> 00:03:00.439 baked into that architecture. Yeah, there is. 00:03:00.500 --> 00:03:02.939 It's the quadratic resource limit. It sounds 00:03:02.939 --> 00:03:05.240 technical, but the idea is because the model 00:03:05.240 --> 00:03:07.280 has to calculate the relationship between every 00:03:07.280 --> 00:03:09.240 word and every single other word. If you double 00:03:09.240 --> 00:03:11.659 the length of the text you feed it, The computation 00:03:11.659 --> 00:03:14.560 cost doesn't just double. It squares. It grows 00:03:14.560 --> 00:03:17.240 exponentially faster. Right. That term quadratic 00:03:17.240 --> 00:03:20.939 growth sounds academic. But when I paste a really 00:03:20.939 --> 00:03:23.580 long article into a chat bot and it slows way 00:03:23.580 --> 00:03:26.259 down, or maybe it just says too long, that's 00:03:26.259 --> 00:03:29.020 the quadratic limit hitting me. Yes. That's exactly 00:03:29.020 --> 00:03:31.099 it. It limits how much text the models can handle 00:03:31.099 --> 00:03:34.180 at once, creating that context window. OK. So 00:03:34.180 --> 00:03:36.419 the transformer could handle these complex connections. 00:03:36.879 --> 00:03:39.580 That ability led directly to the next major shift. 00:03:39.789 --> 00:03:42.669 just scaling things up. In 2020, researchers 00:03:42.669 --> 00:03:45.069 showed that simply making these transformer models 00:03:45.069 --> 00:03:48.569 really, really big thing GPT -3 unlocked this 00:03:48.569 --> 00:03:51.169 completely new skill. It's called few shot learning. 00:03:51.289 --> 00:03:53.310 Few shot learning. This feels like the moment 00:03:53.310 --> 00:03:56.169 AI stopped being just this niche engineering 00:03:56.169 --> 00:03:58.229 thing and started becoming usable for, well, 00:03:58.409 --> 00:04:00.610 almost anyone. Precisely. That's a great way 00:04:00.610 --> 00:04:03.409 to put it. before this huge scaling push. If 00:04:03.409 --> 00:04:06.150 you wanted an AI to do a new task, like summarizing 00:04:06.150 --> 00:04:08.810 customer feedback in a specific way, you needed 00:04:08.810 --> 00:04:12.250 a team of engineers, probably months of GPU compute 00:04:12.250 --> 00:04:15.430 time, training it with thousands, maybe tens 00:04:15.430 --> 00:04:17.930 of thousands of examples. But few shot changed 00:04:17.930 --> 00:04:21.470 that. Why did just making the model bigger suddenly 00:04:21.470 --> 00:04:24.089 enable this? Well, when the models got massive, 00:04:24.269 --> 00:04:26.250 they developed this thing called in -context 00:04:26.250 --> 00:04:28.389 learning. They weren't just predicting the next 00:04:28.389 --> 00:04:30.889 word anymore. They'd seen so many patterns in 00:04:30.889 --> 00:04:32.689 the training data that they actually learned 00:04:32.689 --> 00:04:36.470 to follow specific instructions given in plain 00:04:36.470 --> 00:04:40.490 English. It completely shifted the paradigm from 00:04:40.490 --> 00:04:42.629 training a new model, which is an engineer's 00:04:42.629 --> 00:04:45.550 job, to simply writing a good prompt, which almost 00:04:45.550 --> 00:04:48.250 anyone can do. You just needed one or two good 00:04:48.250 --> 00:04:50.810 examples in the prompt itself. Show the model. 00:04:50.970 --> 00:04:54.529 Product, widget, price, and $10 once. And it 00:04:54.529 --> 00:04:56.310 suddenly knows how to pull the price out of 1 00:04:56.310 --> 00:04:58.589 ,000 other descriptions. That democratization, 00:04:58.790 --> 00:05:00.790 that's really profound. It is. It really is. 00:05:01.290 --> 00:05:04.050 But these early giant models, they were still 00:05:04.050 --> 00:05:06.230 pretty flawed. They were smart. Yeah. But also 00:05:06.230 --> 00:05:08.689 incredibly stubborn sometimes. Excellent at predicting 00:05:08.689 --> 00:05:11.279 the next statistically likely word. but they 00:05:11.279 --> 00:05:14.000 didn't always grasp human intention. They could 00:05:14.000 --> 00:05:17.399 hallucinate very confidently or give answers 00:05:17.399 --> 00:05:19.660 that were just wildly inappropriate or unhelpful. 00:05:19.839 --> 00:05:21.920 Yeah, I still wrestle with prompt drift myself 00:05:21.920 --> 00:05:23.560 sometimes trying to get the output just right. 00:05:23.579 --> 00:05:25.500 Even with the latest models, it's a real thing. 00:05:26.139 --> 00:05:28.879 So this need for helpfulness brought us to the 00:05:28.879 --> 00:05:34.040 next key idea, alignment. Specifically, reinforcement 00:05:34.040 --> 00:05:38.660 learning from human feedback or RLHF. RLHF. This 00:05:38.660 --> 00:05:40.899 is basically the secret sauce that taught the 00:05:40.899 --> 00:05:43.920 AI to be a helpful assistant, not just a text 00:05:43.920 --> 00:05:47.279 generator. They train the AI based on what text 00:05:47.279 --> 00:05:49.660 humans actually preferred. How does that work? 00:05:49.660 --> 00:05:51.680 Like, in practice? It's basically a three -step 00:05:51.680 --> 00:05:54.199 process. First, you have human contractors actually 00:05:54.199 --> 00:05:56.579 write out high -quality, good answers to prompts. 00:05:56.600 --> 00:05:58.519 That's called supervised fine -tuning. Gives 00:05:58.519 --> 00:06:00.660 the model a baseline. OK, step one. Step two, 00:06:00.660 --> 00:06:03.079 they train a separate, smaller model. Its only 00:06:03.079 --> 00:06:05.120 job is to predict which of two answers humans 00:06:05.120 --> 00:06:07.550 would prefer. This is the reward model. Like 00:06:07.550 --> 00:06:10.069 a judge scoring the answers based on human taste. 00:06:10.490 --> 00:06:12.850 Exactly, like a high score for helpfulness. Then 00:06:12.850 --> 00:06:14.629 the final step is the reinforcement learning 00:06:14.629 --> 00:06:17.550 part. They let the main AI generate answers, 00:06:17.910 --> 00:06:21.329 the reward model scores them instantly, and the 00:06:21.329 --> 00:06:24.069 AI adjusts its own parameters to try and maximize 00:06:24.069 --> 00:06:26.170 that reward score. It's like training a dog with 00:06:26.170 --> 00:06:28.230 treats basically, reinforcing the good behavior. 00:06:28.399 --> 00:06:30.779 And the big insight there was that a smally model 00:06:30.779 --> 00:06:33.199 that was aligned and struck GPT was actually 00:06:33.199 --> 00:06:36.160 preferred by users much more than the giant but 00:06:36.160 --> 00:06:39.420 unaligned GPT -3. Usefulness beats sheer size. 00:06:40.079 --> 00:06:41.920 Yes. Usefulness suddenly became the key metric. 00:06:42.339 --> 00:06:45.240 Alignment was critical. So if alignment's so 00:06:45.240 --> 00:06:49.000 crucial and RLHF was the way, why are we seeing 00:06:49.000 --> 00:06:51.600 newer, maybe simpler methods starting to replace 00:06:51.600 --> 00:06:54.800 it now? Well, RLHF is quite complex and, frankly, 00:06:54.939 --> 00:06:57.189 very expensive to implement. that's driving research 00:06:57.189 --> 00:06:59.310 into simpler, cheaper alignment methods like 00:06:59.310 --> 00:07:01.889 DPO. Okay, so we have this aligned AI brain. 00:07:01.949 --> 00:07:03.529 Now let's talk about connecting it to the real 00:07:03.529 --> 00:07:06.970 world. First up is RAG, retrieval augmented generation. 00:07:07.430 --> 00:07:09.350 This is essentially giving the AI an outside 00:07:09.350 --> 00:07:11.730 brain that can access current information. Right, 00:07:11.829 --> 00:07:14.470 because we built this giant LLM brain, but its 00:07:14.470 --> 00:07:16.490 knowledge is frozen at the time it was trained. 00:07:16.699 --> 00:07:19.339 Why couldn't we just retrain it more often to 00:07:19.339 --> 00:07:21.620 keep it updated? Because retraining one of these 00:07:21.620 --> 00:07:25.139 massive models costs potentially tens of millions 00:07:25.139 --> 00:07:27.759 of dollars and can take weeks or months. It's 00:07:27.759 --> 00:07:30.279 just not feasible to do it constantly. So its 00:07:30.279 --> 00:07:34.560 knowledge gets stale fast. Margay solves this. 00:07:34.879 --> 00:07:38.000 It works by first finding relevant external info, 00:07:38.459 --> 00:07:40.620 maybe real -time news, maybe private company 00:07:40.620 --> 00:07:44.220 documents. Then it adds that specific text directly 00:07:44.220 --> 00:07:47.209 into the prompt it sends to the AI. OK. And crucially, 00:07:47.470 --> 00:07:50.370 it forces the AI to generate its answer based 00:07:50.370 --> 00:07:52.750 only on that source text provided in the prompt. 00:07:53.089 --> 00:07:55.649 So if I ask my bank's chatbot about, say, my 00:07:55.649 --> 00:07:57.829 specific mortgage rate, which is private info. 00:07:58.050 --> 00:07:59.990 Argue would search the bank's secure document 00:07:59.990 --> 00:08:02.189 database, find the paragraph with your rate, 00:08:02.610 --> 00:08:04.589 paste only that paragraph into the prompt for 00:08:04.589 --> 00:08:07.050 the LLM, and instruct it, answer the customer 00:08:07.050 --> 00:08:09.730 using only this text. Got it. So the main LLM 00:08:09.730 --> 00:08:12.050 never actually gets trained on or learns my private 00:08:12.050 --> 00:08:14.560 data. It just uses it for that one answer. Exactly. 00:08:14.639 --> 00:08:17.800 That protects privacy and it also massively reduces 00:08:17.800 --> 00:08:20.439 hallucinations because the AI is grounded in 00:08:20.439 --> 00:08:22.699 a specific source document. Makes sense. What's 00:08:22.699 --> 00:08:25.100 the main risk then when you're relying on a ROJ 00:08:25.100 --> 00:08:27.620 system? Well, the final answer quality depends 00:08:27.620 --> 00:08:29.899 entirely on that first step, the retrieval or 00:08:29.899 --> 00:08:32.820 search step. If the search pulls up bad or irrelevant 00:08:32.820 --> 00:08:35.980 info, the AI's answer will be bad too. Garbage 00:08:35.980 --> 00:08:38.820 in, garbage out. Okay. That brings us to the 00:08:38.820 --> 00:08:41.980 next step. Agents. This feels like a really big 00:08:41.980 --> 00:08:45.340 shift, moving from the AI being a passive chat 00:08:45.340 --> 00:08:48.000 bot waiting for me to type something to being 00:08:48.000 --> 00:08:50.320 an active tool that can actually go out and do 00:08:50.320 --> 00:08:52.580 things to achieve a goal. That's exactly it. 00:08:52.820 --> 00:08:55.299 Agents are about planning, using tools like running 00:08:55.299 --> 00:08:58.120 a web search, executing code, calling an external 00:08:58.120 --> 00:09:01.580 API like weather or stocks, and then, importantly, 00:09:01.909 --> 00:09:04.009 observing the results and correcting mistakes. 00:09:04.470 --> 00:09:06.730 So what's the structure? How does an agent work? 00:09:06.870 --> 00:09:09.590 It's pretty simple conceptually. You have a brain, 00:09:09.809 --> 00:09:11.590 which is usually the LLM doing the high level 00:09:11.590 --> 00:09:13.870 thinking and planning. You have perception, which 00:09:13.870 --> 00:09:15.649 is the agency and the results of the tools it 00:09:15.649 --> 00:09:18.309 uses, and action, which is actually using those 00:09:18.309 --> 00:09:21.889 tools. And it works in this loop. Think, act, 00:09:22.690 --> 00:09:26.649 act, see the result, think again based on the 00:09:26.649 --> 00:09:30.190 result. Over and over. until the goal is met. 00:09:30.490 --> 00:09:32.690 So instead of me asking like three separate questions, 00:09:32.809 --> 00:09:34.490 what's the weather in Hanoi? What's the weather 00:09:34.490 --> 00:09:36.289 in Ho Chi Minh City? OK, based on that, what 00:09:36.289 --> 00:09:38.690 should I pack? Right. I could just give the agent 00:09:38.690 --> 00:09:41.950 one complex goal, like compare the weather forecast 00:09:41.950 --> 00:09:44.649 for Hanoi and HCMC for the next three days and 00:09:44.649 --> 00:09:46.830 suggest what clothes I should pack for a business 00:09:46.830 --> 00:09:49.070 trip. Exactly that. You give it the complex goal. 00:09:49.690 --> 00:09:51.769 Analyze the last three financial reports for 00:09:51.769 --> 00:09:54.149 Company X, check recent market sentiment about 00:09:54.149 --> 00:09:56.590 them on Twitter, and draft me a summary email 00:09:56.590 --> 00:09:58.529 recommending whether I should buy or sell the 00:09:58.529 --> 00:10:01.389 stock. The agent figures out the steps and uses 00:10:01.389 --> 00:10:05.200 its tools sequentially. Whoa! Okay, imagine scaling 00:10:05.200 --> 00:10:08.480 that agent structure up to manage, say, a billion 00:10:08.480 --> 00:10:11.000 dynamic calendar scheduling requests a day across 00:10:11.000 --> 00:10:13.600 a huge company. That changes everything about 00:10:13.600 --> 00:10:16.360 how work gets done almost instantly. That's the 00:10:16.360 --> 00:10:18.279 potential, absolutely. It's the immediate future 00:10:18.279 --> 00:10:21.460 of productivity enhancement. But these systems 00:10:21.460 --> 00:10:23.700 are still pretty new and can be tricky to manage 00:10:23.700 --> 00:10:26.220 reliably. Right, they are complex. So what's 00:10:26.220 --> 00:10:29.120 the biggest, like, operational headache or risk 00:10:29.120 --> 00:10:31.720 when people try to deploy agents in the real 00:10:31.720 --> 00:10:34.220 world today? They can still get stuck sometimes. 00:10:34.500 --> 00:10:37.159 They might get into self -repeating loops, like 00:10:37.159 --> 00:10:39.320 endlessly searching for a file that doesn't exist 00:10:39.320 --> 00:10:42.320 or calling a broken tool over and over. Reliability 00:10:42.320 --> 00:10:45.059 is still a challenge. So, okay, we've built this 00:10:45.059 --> 00:10:48.840 powerful, aligned, goal -seeking AI. Awesome. 00:10:49.440 --> 00:10:51.679 But for a while, it remained way too huge and 00:10:51.679 --> 00:10:53.919 expensive for most individuals or smaller companies 00:10:53.919 --> 00:10:56.120 to actually run themselves. Right. Locked up 00:10:56.120 --> 00:10:58.320 in big tech clouds mostly. Exactly. So the next 00:10:58.320 --> 00:11:00.500 three concepts we need to touch on really solve 00:11:00.500 --> 00:11:02.980 this accessibility and cost problem. made it 00:11:02.980 --> 00:11:04.919 more democratic. This is what some people call 00:11:04.919 --> 00:11:08.919 the efficiency triad. LoRa, MoE, and quantization, 00:11:09.399 --> 00:11:11.980 basically making giant AI cheaper, faster, and 00:11:11.980 --> 00:11:14.639 much easier to deploy. Let's start with LoRa 00:11:14.639 --> 00:11:17.320 low -rank adaptation. OK, LoRa, think of the 00:11:17.320 --> 00:11:20.960 massive bass AI model as like a giant expertly 00:11:20.960 --> 00:11:23.980 pre -trained symphony orchestra. Okay, orchestra. 00:11:24.259 --> 00:11:26.240 Now, if you wanted that orchestra to learn a 00:11:26.240 --> 00:11:28.779 completely new style, say experimental jazz, 00:11:29.179 --> 00:11:32.059 the old way full fine -tuning was like retraining 00:11:32.059 --> 00:11:34.240 every single musician on every single instrument. 00:11:34.580 --> 00:11:37.860 Hugely expensive. Massive amounts of data, time, 00:11:38.360 --> 00:11:40.519 storage. Prohibitively expensive, yeah. And you'd 00:11:40.519 --> 00:11:43.139 end up with a whole new giant orchestra file. 00:11:43.879 --> 00:11:46.419 Laura completely sidesteps this. Laura says, 00:11:46.820 --> 00:11:49.259 keep 99 % of the original orchestra musicians 00:11:49.259 --> 00:11:52.909 frozen. Don't touch them. Just add a few small 00:11:52.909 --> 00:11:55.990 new specialized pieces. Think like adding a dedicated 00:11:55.990 --> 00:11:58.570 jazz conductor and maybe a specific drummer. 00:11:59.289 --> 00:12:02.309 Then you only train those tiny new adapter layers 00:12:02.309 --> 00:12:04.649 to learn the jazz style. Ah, so you end up with 00:12:04.649 --> 00:12:07.389 just the original massive model plus this tiny 00:12:07.389 --> 00:12:09.090 little instruction file that tells it how to 00:12:09.090 --> 00:12:11.850 play jazz when needed. Precisely. It allows you 00:12:11.850 --> 00:12:14.190 to fine -tune a huge model, often using just 00:12:14.190 --> 00:12:17.190 a single consumer GPU, and the resulting adapter 00:12:17.190 --> 00:12:20.210 file might only be, say, 100 megabytes instead 00:12:20.210 --> 00:12:22.009 of hundreds of gigabytes. That's why the open 00:12:22.009 --> 00:12:24.029 -source community exploded with custom models, 00:12:24.230 --> 00:12:26.149 right? Yeah. Specialization became cheap and 00:12:26.149 --> 00:12:30.269 portable. Totally. Okay, next up, Moe, mixture 00:12:30.269 --> 00:12:33.750 of experts, popularized by models like Google's 00:12:33.750 --> 00:12:36.620 Switch Transformer and Mixeroll. This tackles 00:12:36.620 --> 00:12:39.500 the speed problem of running these enormous models. 00:12:39.879 --> 00:12:42.299 Right. How can a model with maybe a trillion 00:12:42.299 --> 00:12:45.960 parameters run fast? It's a really clever architectural 00:12:45.960 --> 00:12:49.080 trick. Imagine the AI model is now a massive 00:12:49.080 --> 00:12:52.419 hospital staffed with like a thousand different 00:12:52.419 --> 00:12:54.440 medical specialists. Okay, hospital analogy. 00:12:54.600 --> 00:12:57.399 In the old dense model architecture, every time 00:12:57.399 --> 00:12:59.600 any patient came in, even with just a common 00:12:59.600 --> 00:13:03.240 cold, All 1 ,000 specialist doctors had to consult 00:13:03.240 --> 00:13:05.759 on the case. A huge waste of expert time. Right. 00:13:05.980 --> 00:13:07.940 Makes sense. The MOE approach is different. Yeah. 00:13:08.299 --> 00:13:09.919 With MOE, there's a quick router at the front 00:13:09.919 --> 00:13:12.720 desk. When a patient or a query comes in talking 00:13:12.720 --> 00:13:15.580 about, say, programming, the router sends them 00:13:15.580 --> 00:13:17.860 only to the programming expert wing of the hospital. 00:13:18.379 --> 00:13:20.639 Only that relevant small set of specialists gets 00:13:20.639 --> 00:13:23.960 activated. Ah. So. Companies can build these 00:13:23.960 --> 00:13:26.320 models with trillions of parameters, making them 00:13:26.320 --> 00:13:28.480 incredibly knowledgeable across many domains. 00:13:28.539 --> 00:13:31.139 Yes. But for any single question, they only actually 00:13:31.139 --> 00:13:33.460 run a small fraction of those parameters. Maybe 00:13:33.460 --> 00:13:36.220 just a relevant expert. That's exactly it. So 00:13:36.220 --> 00:13:38.840 we went from building one massive general practitioner 00:13:38.840 --> 00:13:40.860 brain that had to read every textbook for every 00:13:40.860 --> 00:13:43.759 patient to building a huge team of specialists, 00:13:44.259 --> 00:13:46.320 but only calling in the one needed for the job. 00:13:46.659 --> 00:13:48.600 That's how they stay fast, despite the enormous 00:13:48.600 --> 00:13:51.159 total size. Clever. OK, and the third part of 00:13:51.159 --> 00:13:54.450 the efficiency triad. Quantization. Quantization, 00:13:54.610 --> 00:13:56.649 the memory hack. This is basically saving memory 00:13:56.649 --> 00:13:59.350 by using less precise numbers, like rounding. 00:13:59.450 --> 00:14:01.710 Pretty much. It's a pure engineering optimization. 00:14:02.490 --> 00:14:04.889 AI model weights, the parameters, are often stored 00:14:04.889 --> 00:14:06.889 as very precise numbers, like 16 -bit floating 00:14:06.889 --> 00:14:09.950 point numbers. Quantization is like saying, OK, 00:14:10.090 --> 00:14:14.990 instead of storing pi as 3 .14159265, let's just 00:14:14.990 --> 00:14:18.230 store it as 3 .14. It's good enough for the calculation. 00:14:18.350 --> 00:14:21.519 Often, yes. For many models, reducing the precision 00:14:21.519 --> 00:14:23.460 maybe down to 8 -bit integers, Intellidate, or 00:14:23.460 --> 00:14:25.919 even 4 -bit cuts the memory requirement roughly 00:14:25.919 --> 00:14:28.720 in half, or even more, without a major drop in 00:14:28.720 --> 00:14:30.879 performance quality. And this trick, this is 00:14:30.879 --> 00:14:33.740 what allows huge models like Metos Llama 3 to 00:14:33.740 --> 00:14:36.080 potentially run not just on giant server farms, 00:14:36.240 --> 00:14:38.759 but maybe on a high -end gaming PC, or eventually 00:14:38.759 --> 00:14:41.929 even your smartphone. That's the goal. It bridges 00:14:41.929 --> 00:14:45.649 the gap between AI being purely a cloud or corporate 00:14:45.649 --> 00:14:48.909 asset and becoming a truly personal, locally 00:14:48.909 --> 00:14:51.649 runnable tool. Huge implications. Okay, so we 00:14:51.649 --> 00:14:53.509 have the engine, it's efficient, it's connected. 00:14:53.809 --> 00:14:56.330 What's the last piece? The last really critical 00:14:56.330 --> 00:14:59.950 concept addresses the final hurdle for agents 00:14:59.950 --> 00:15:02.110 to really take off and work together seamlessly. 00:15:02.710 --> 00:15:05.070 The need for a common language or standard. Right, 00:15:05.169 --> 00:15:08.330 the model context protocol or MCP. Yeah, MCP. 00:15:08.600 --> 00:15:11.700 This aims to solve what developers call the N 00:15:11.700 --> 00:15:14.200 by M problem. Yeah. Imagine you have a hundred 00:15:14.200 --> 00:15:16.960 different AI models or agents. Okay. And you 00:15:16.960 --> 00:15:18.700 have maybe a thousand different digital tools 00:15:18.700 --> 00:15:20.879 you want them to use. Notion, Slack, Google Calendar, 00:15:21.120 --> 00:15:22.779 Salesforce, whatever. Right now you'd have to 00:15:22.779 --> 00:15:25.179 write custom code, like specific glue, to connect 00:15:25.179 --> 00:15:27.940 every single model to every single tool. That's 00:15:27.940 --> 00:15:29.820 a hundred times a thousand, hundred thousand 00:15:29.820 --> 00:15:32.120 custom connections. A completely unsustainable 00:15:32.120 --> 00:15:34.960 integration nightmare. Exactly. MCP wants to 00:15:34.960 --> 00:15:37.559 be like the universal USB standard for AI tools. 00:15:37.929 --> 00:15:40.509 unplugged to rule them all. Kind of. Tool developers 00:15:40.509 --> 00:15:44.289 just implement one standard MCP server interface 00:15:44.289 --> 00:15:47.549 for their tool. Then any AI agent that understands 00:15:47.549 --> 00:15:50.409 the MCP standard can instantly plug in and use 00:15:50.409 --> 00:15:53.350 that tool. No custom code needed. That seems 00:15:53.350 --> 00:15:55.929 absolutely critical if we want agents to eventually 00:15:55.929 --> 00:15:58.509 manage our whole digital life smoothly. It's 00:15:58.509 --> 00:16:00.429 fundamental for realizing the true potential 00:16:00.429 --> 00:16:04.830 of interconnected AI agents. So let's just quickly 00:16:04.830 --> 00:16:06.950 recap the journey we took through these foundational 00:16:06.950 --> 00:16:09.139 concepts from the source material. It's quite 00:16:09.139 --> 00:16:11.620 a story. We started with the core engine, the 00:16:11.620 --> 00:16:14.019 breakthrough idea of attention and the transformer 00:16:14.019 --> 00:16:16.059 architecture. Yep, built the engine. Then we 00:16:16.059 --> 00:16:18.320 scaled that engine way up with models like GPT 00:16:18.320 --> 00:16:20.679 -3, unlocking few -shot learning, the ability 00:16:20.679 --> 00:16:23.120 to instruct AI with prompts, not just program 00:16:23.120 --> 00:16:25.200 it with data. Right. Then we realized raw power 00:16:25.200 --> 00:16:27.440 wasn't enough. We needed alignment. We taught 00:16:27.440 --> 00:16:30.200 the AI to be helpful and safe using human preferences 00:16:30.200 --> 00:16:33.000 through techniques like RLHF. Made it useful. 00:16:33.230 --> 00:16:36.169 And then we connected that aligned brain to the 00:16:36.169 --> 00:16:39.309 live dynamic world, using RG to give it access 00:16:39.309 --> 00:16:42.250 to real -time data safely, and empowering it 00:16:42.250 --> 00:16:45.350 to actually act on goals using the agent framework. 00:16:45.950 --> 00:16:49.169 And finally, the efficiency revolution, making 00:16:49.169 --> 00:16:52.570 all this incredible power accessible, affordable, 00:16:52.769 --> 00:16:55.509 and fast enough for widespread use through clever 00:16:55.509 --> 00:16:58.710 engineering like LoRa, MoE, and quantization. 00:16:58.830 --> 00:17:01.899 Made it practical. Yeah. The field moves incredibly 00:17:01.899 --> 00:17:04.380 fast, feels like it sometimes, but when you break 00:17:04.380 --> 00:17:06.960 it down like this, these core building blocks, 00:17:07.160 --> 00:17:09.500 they're actually quite understandable. You really 00:17:09.500 --> 00:17:12.339 are. And what's truly fascinating now... building 00:17:12.339 --> 00:17:15.200 on that last point about MCP, the common language, 00:17:15.779 --> 00:17:17.819 is that the next big frontier might not just 00:17:17.819 --> 00:17:20.200 be about smarter algorithms or bigger models. 00:17:20.660 --> 00:17:22.980 It's shifting towards better governance and better 00:17:22.980 --> 00:17:26.180 standards. Right. That MCP idea raises a huge 00:17:26.180 --> 00:17:28.200 question for the future, doesn't it? It really 00:17:28.200 --> 00:17:31.660 does. As these agents using protocols like MCP 00:17:31.660 --> 00:17:33.819 gain the ability to plug into and potentially 00:17:33.819 --> 00:17:35.880 manage everything, your email, your bank accounts, 00:17:36.019 --> 00:17:38.160 your work calendar, your smart home, how are 00:17:38.160 --> 00:17:41.339 we as a society going to solve the immense security 00:17:41.339 --> 00:17:44.339 challenges, the liability questions, the need 00:17:44.339 --> 00:17:46.460 for industry consensus to make this safe and 00:17:46.460 --> 00:17:48.480 reliable for mass adoption. That's the big one 00:17:48.480 --> 00:17:51.279 to think about. That's really the challenge for 00:17:51.279 --> 00:17:53.880 all of us, for you listening, to consider as 00:17:53.880 --> 00:17:56.579 you watch this space evolve incredibly rapidly 00:17:56.579 --> 00:17:58.119 over the next few years. It's going to be quite 00:17:58.119 --> 00:18:01.180 a ride. Thanks for diving deep into the sources 00:18:01.180 --> 00:18:01.720 with us today.