WEBVTT 00:00.440 --> 00:03.680 Today we are going to talk about large language models. 00:05.440 --> 00:12.960 Before we start cooking and diving into all these exciting topics like powerful AI agents, automations 00:12.960 --> 00:19.400 that can transform your work and business, let's go back to the basics and briefly discuss large language 00:19.400 --> 00:19.960 models. 00:21.760 --> 00:22.160 Why? 00:22.200 --> 00:28.720 Because understanding how they work will help you create more advanced and powerful AI agents. 00:28.760 --> 00:30.880 Automation and AI systems. 00:32.040 --> 00:36.640 This knowledge is going to be crucial as we move into more advanced topics. 00:38.360 --> 00:41.200 Alright, so let's start with a simple question. 00:41.640 --> 00:44.120 What actually is a large language model? 00:44.760 --> 00:51.440 If you have used the ChatGPT Cloud Llama two or any other modern AI assistant. 00:52.320 --> 00:54.240 But let's take a step back. 00:55.160 --> 00:57.240 What do these models actually do? 00:57.800 --> 01:05.250 As decor, they They're just really, really good at predicting the next word in a sentence. 01:06.570 --> 01:12.610 Imagine you are texting someone and your phone suggests words to complete your sentence. 01:12.850 --> 01:15.290 Something like hey, how are you? 01:16.330 --> 01:19.690 You expect the word doing or today to appear? 01:20.250 --> 01:27.690 That's because your phone has seen common phrases and makes predictions based on what it has learned. 01:28.370 --> 01:32.330 Llms work the same way, but on a much larger scale. 01:33.730 --> 01:40.810 They've had to read massive amounts of text like books, articles, code, conversations, and use that 01:40.810 --> 01:44.770 knowledge to make incredibly accurate predictions. 01:45.890 --> 01:47.610 They don't think like humans. 01:48.130 --> 01:50.650 They don't understand like like we do. 01:51.490 --> 01:56.170 They just use probabilities to determine the most likely next word. 01:56.730 --> 02:00.070 And when you scale this up with billions of examples? 02:00.430 --> 02:02.870 You get something that feels intelligent. 02:04.870 --> 02:12.630 Now, if we strip away all the marketing buzz and hype, what's an LLM actually made of? 02:14.150 --> 02:19.070 Well, surprisingly, LLM is just two files. 02:19.390 --> 02:26.150 The first one is a parameters file which contains all the learned knowledge. 02:27.070 --> 02:32.150 And the second file is a code file which runs the model to generate text. 02:33.430 --> 02:36.150 Now let's compare it to a music player. 02:36.430 --> 02:38.710 So imagine you have a playlist of songs. 02:39.430 --> 02:47.110 And that playlist is like the parameters file because it stores all the information, the lyrics, the 02:47.110 --> 02:47.870 melodies. 02:48.750 --> 02:51.830 And now you need something to actually play the music. 02:53.230 --> 03:00.880 That's where the code file comes in, because it's the music player that reads the playlist and makes 03:00.880 --> 03:04.240 sound for lmms. 03:04.280 --> 03:12.080 The parameters file contains billions of numbers, each representing what the model has learned. 03:12.960 --> 03:17.320 So it's a compressed version of all the text it has processed. 03:18.080 --> 03:25.480 The code file is a relatively small program that takes in a prompt, applies mathematical functions, 03:26.360 --> 03:31.320 and then spits out the the next word based on probability. 03:34.440 --> 03:38.240 Now how does an LLM actually learn? 03:38.560 --> 03:42.960 How do you go from raw text to a working AI model? 03:43.440 --> 03:48.520 So the answer is a lot of data and a ridiculous amount of computing power. 03:49.880 --> 03:53.160 So first companies collect massive datasets. 03:54.240 --> 04:02.490 Think of like books, Wikipedia, news articles, scientific papers, Reddit discussions, even public 04:02.490 --> 04:03.690 code repositories. 04:04.010 --> 04:08.370 And the goal is to feed the model as much human knowledge as possible. 04:10.970 --> 04:12.370 Next comes training. 04:12.570 --> 04:17.330 So this is where the magic happens because the model starts with random numbers. 04:18.410 --> 04:25.850 Yes, actually at the start it knows nothing and we force it to predict the next word in billions of 04:25.850 --> 04:26.770 sentences. 04:28.050 --> 04:34.850 When it gets a word wrong, we update the parameters using a technique called gradient descent. 04:35.290 --> 04:37.490 And over time it gets better. 04:37.970 --> 04:47.090 So this process runs for weeks on thousands of GPUs, which are the super fast processors used for AI. 04:47.130 --> 04:55.140 Training is kind of like taking all the knowledge of the internet and squishing it down into a compressed 04:55.140 --> 04:57.660 format that fits in a single file. 05:00.020 --> 05:03.980 Okay, now let's go deeper into how these models actually generate text. 05:05.100 --> 05:07.660 So it all comes down to one simple idea. 05:08.300 --> 05:09.820 Predicting the next word. 05:10.860 --> 05:12.300 So let's take an example. 05:13.540 --> 05:15.780 The cat sat on a chair. 05:16.780 --> 05:19.180 And what word do you expect to come next. 05:19.460 --> 05:21.980 Most people would say mat. 05:22.660 --> 05:23.140 But why? 05:23.180 --> 05:26.860 Because you've seen the phrase before and it's very common. 05:27.940 --> 05:30.020 Neelam does the same thing. 05:30.340 --> 05:33.020 It looks at the words before the blank. 05:35.460 --> 05:39.340 And assigns probabilities to possible next words. 05:40.340 --> 05:43.740 So the floor is about 10% probability. 05:43.940 --> 05:47.220 So for 5% and 31%. 05:48.940 --> 05:56.720 So the model picks the most probable one and moves on to predict the next word and the next and the 05:56.720 --> 05:57.200 next. 05:57.960 --> 06:05.280 So this simple mechanism, repeated billions of times, leads to intelligent sounding conversations. 06:06.000 --> 06:09.480 Now we need to differentiate two types of models. 06:09.600 --> 06:11.600 Open source models and closed models. 06:12.600 --> 06:15.720 Because not all llms are the same. 06:16.360 --> 06:19.480 Some are open source and some are closed source. 06:20.160 --> 06:28.040 So open source models like Deepsea Llama two, Mistral Falcon are free to download and modify so you 06:28.040 --> 06:29.920 can run them on your own computer. 06:30.520 --> 06:34.080 Tweak them and fine tune them for specific tasks. 06:36.520 --> 06:45.560 On the other hand, closed source models like GPT four, Gemini Cloud are owned by companies like OpenAI 06:45.920 --> 06:46.640 and Google. 06:47.720 --> 06:54.010 You can use them, but don't get access to the architecture and how they work internally. 06:54.330 --> 06:57.090 So open source models give you more control. 06:58.130 --> 07:02.730 Now, raw llms are actually not that smart. 07:03.090 --> 07:08.290 They can generate text, but they don't naturally follow instructions well. 07:09.050 --> 07:15.210 So to turn them into useful assistants like ChatGPT, they go through a process called fine tuning. 07:16.050 --> 07:26.090 And fine tuning basically means taking a trained LLM and giving it specific examples of how it should 07:26.090 --> 07:27.610 respond to questions. 07:27.770 --> 07:36.930 Humans write sample conversations, and the model is adjusted so that it learns to mimic helpful, structured 07:37.410 --> 07:38.250 responses. 07:39.450 --> 07:41.250 So this is how ChatGPT knows. 07:41.450 --> 07:50.780 To answer questions, politely summarize articles, and explain concepts clearly because it was trained 07:50.780 --> 07:54.340 to do so after the initial training phase. 07:57.820 --> 08:05.660 Now, fine tuning takes a pre-trained LLM and specializes it for a specific task or use case. 08:05.940 --> 08:13.620 So this is done by providing the model with a curated data set, which is basically a collection of 08:14.300 --> 08:22.020 high quality human written examples that teach the model how to respond in a specific way. 08:23.820 --> 08:31.180 So fine tuned LLM that performs better in a targeted domain, whether that's answering customer service 08:31.180 --> 08:34.140 questions or writing code. 08:39.180 --> 08:47.980 Now, every few months, a new, more powerful LLM comes out because improving AI is as simple as adding 08:48.350 --> 08:56.350 more training data and more computing power so each generation of models gets bigger, faster and smarter. 08:56.670 --> 09:05.230 Just like upgrading a car engine, newer models perform better than the older ones, so the cycle will 09:05.230 --> 09:07.750 continue as AI technology evolves. 09:15.430 --> 09:17.550 So AI is no longer just a chatbot. 09:18.230 --> 09:22.990 It's evolving into an operating system for knowledge and automation. 09:23.430 --> 09:30.870 Right now, we are already seeing AI models that don't just generate text, but can search the web, 09:30.910 --> 09:36.510 run code, interact with APIs, and even control applications. 09:37.150 --> 09:44.870 So these functionalities mean AI is moving beyond passive response generation and becoming an active 09:46.360 --> 09:48.920 Kind of participant in workflows. 09:50.760 --> 09:58.440 And this is where things get really interesting, because if I can process text, images, audio and 09:58.440 --> 10:05.680 interact with external tools like systems and apps, what are we actually talking about? 10:07.320 --> 10:08.480 Are I agents? 10:11.280 --> 10:13.240 Are agents are the next step. 10:14.200 --> 10:21.400 These are intelligent systems that don't just generate responses but take actions on their own. 10:21.920 --> 10:28.240 They can schedule tasks, automate workflows, analyze real time data, and even make decisions based 10:28.240 --> 10:29.240 on context. 10:30.160 --> 10:34.200 So now let's shift gears and dive into this next evolution. 10:36.240 --> 10:37.120 Agent. 10:37.960 --> 10:43.120 So we'll go through how they work and how they are shaping the future of automation. 10:47.300 --> 10:49.620 Let's quickly recap what we have covered today. 10:50.180 --> 10:56.060 So we start with LMS, which predicts words based on vast amounts of training data. 10:56.500 --> 11:03.100 Then we explored how fine tuning transfers these raw models into useful AI systems. 11:04.740 --> 11:06.820 But AI is evolving beyond just text. 11:06.820 --> 11:14.620 Generation is becoming multi-modal and understanding not just words, but images, audio, and real 11:14.620 --> 11:15.420 world data. 11:17.500 --> 11:24.740 And this shift is turning AI into an operating system for automation, where models don't just answer 11:24.740 --> 11:28.260 questions but interact with apps, tools, and the internet. 11:29.860 --> 11:38.460 And now we are entering the era of agents for systems that don't just generate responses, but actually 11:38.500 --> 11:40.420 take actions on their own. 11:42.180 --> 11:44.620 Thanks for your attention and see you in the next lesson.