WEBVTT 00:01.320 --> 00:03.080 Hello everyone and welcome. 00:03.400 --> 00:07.600 In this video we will learn about evolution of large language models. 00:07.920 --> 00:13.560 We'll learn how large language models evolve from being a model that helps with prediction, to adding 00:13.600 --> 00:19.360 technology advancements to the model, and how it can do more than just prediction. 00:19.800 --> 00:23.560 So let's dive deeper into evolution of large language models. 00:23.760 --> 00:29.440 So what are the LLM capabilities, limitations, and path to MCP? 00:30.080 --> 00:39.240 Large language models like ChatGPT represent a significant breakthrough in early artificial intelligence. 00:39.520 --> 00:46.800 These systems excel at numerous complex tasks like including code generation, text summarization, 00:47.000 --> 00:48.880 and even image creation. 00:49.440 --> 00:54.840 There were centrality has made them valuable across many industries and applications. 00:56.480 --> 01:00.600 However, this technology had some initial limitation. 01:00.600 --> 01:06.200 Despite their impressive capabilities, there are notable constraints on Lem. 01:06.720 --> 01:09.840 Consider the first version of ChatGPT. 01:10.560 --> 01:16.840 If you asked it to send an email on your behalf, it would simply respond saying, I cannot perform 01:16.840 --> 01:17.760 that action. 01:18.320 --> 01:23.800 This highlights a fundamental limitation of traditional large language model. 01:24.480 --> 01:27.080 What are the core functions of Lem? 01:27.560 --> 01:30.200 It's a sophisticated prediction engine. 01:30.680 --> 01:37.640 They analyze patterns in the text and predict what word or phrase should come next in the sequence. 01:38.080 --> 01:41.160 So, for example, if you say I want to play. 01:41.600 --> 01:49.280 The model uses its training data to determine what is the most likely next ward going to be. 01:49.800 --> 01:54.040 It can be games, music, or anything that it is trained on. 01:54.480 --> 01:59.640 So the next token prediction is essentially the mechanism that powers all llms. 01:59.640 --> 01:59.750 LMS. 02:00.150 --> 02:01.750 So why does it matter? 02:02.070 --> 02:11.190 Recognizing this, limitations help us understand why technologies like Model Context Protocol presents 02:11.190 --> 02:13.870 the next evolutionary step for LM. 02:14.430 --> 02:22.070 While LM excels at understanding language generation, they needed additional capabilities to interact 02:22.070 --> 02:26.230 with external systems and perform real world tasks. 02:26.750 --> 02:31.710 So now let's go over the first evolution of LM, which was. 02:32.950 --> 02:36.150 That is retrieval augmented generation. 02:36.390 --> 02:43.510 It's a technique that enhances large language models by giving them access to external information or 02:43.550 --> 02:46.030 sources that it is not trained on. 02:46.430 --> 02:51.990 The core problem that Rags solves is llms are trained on certain specific data. 02:52.390 --> 02:59.590 They cannot use information that is outside the training data to analyze and answer your questions, 03:00.630 --> 03:06.390 something like whether or company specific data is something that it's not trained on. 03:06.790 --> 03:13.630 So this creates a significant limitations on the accurate and up to date specialized information that 03:14.510 --> 03:15.270 can help. 03:15.750 --> 03:17.350 So how Rag works. 03:17.470 --> 03:20.190 It has the first step of query processing. 03:20.510 --> 03:26.230 When you ask question the system, analyze what information might be needed. 03:26.550 --> 03:28.630 Then we have information retrieval. 03:29.110 --> 03:34.110 It searches through external database documents or knowledge sources. 03:34.590 --> 03:36.870 Then there is context integration. 03:37.270 --> 03:46.830 The retrieved information is fed to the along with the original question, and then generates an answer 03:46.830 --> 03:51.990 using both its trading knowledge and the information that was retrieved. 03:52.510 --> 04:00.580 So I have covered Rag in a different section where we would take a deep dive into this concept. 04:00.580 --> 04:08.380 However, I briefly went over Rag here so that we understand how the Llms are evolving over time. 04:09.060 --> 04:15.340 So now that we had Rag, there were still challenges and we wanted to extend the LLM capabilities. 04:15.740 --> 04:18.660 The next thing that was born is AI agent. 04:18.980 --> 04:21.220 So what is AI agent? 04:21.700 --> 04:27.540 AI agent is a system that combines large language models with external tools. 04:27.900 --> 04:36.180 You can think of AI agents as system that combines LLM with external tools to perform autonomous and 04:36.180 --> 04:37.620 complex tasks. 04:37.980 --> 04:45.460 This was a significant leap beyond expeditions that LLM were able to work with external entities to 04:45.500 --> 04:49.180 solve more complex problem at an enterprise level. 04:49.620 --> 04:56.660 You can think of agents as tools that can integrate with application programming interface that allows 04:57.420 --> 05:01.260 them Interact with external services and systems. 05:01.620 --> 05:09.420 These tools extend the LMS capability far beyond text generation, enabling them to perform real world 05:09.420 --> 05:12.580 actions and access necessary information. 05:13.140 --> 05:20.940 So now let's take a real world example of LM using agents, and also learn about some of the challenges 05:20.940 --> 05:22.340 that comes with it. 05:22.660 --> 05:26.460 So here is a real world example perplexity. 05:26.980 --> 05:32.740 That's the search engine that defines and demonstrates the evolution very effectively. 05:33.180 --> 05:37.580 So when you chat with perplexity you're not using just the LM. 05:37.860 --> 05:46.740 You are using a system that LM also does internet search to retrieve the information that you need for 05:46.740 --> 05:49.220 the search that you perform on the website. 05:49.700 --> 05:56.260 LL itself cannot browse the internet, but the integration tools make it possible. 05:56.740 --> 05:59.090 There are also challenges in this area. 05:59.090 --> 06:01.130 While tool integration works well. 06:01.610 --> 06:07.170 Building a comprehensive AI assistant becomes increasingly complex. 06:08.970 --> 06:16.650 Imagine creating a system that can search internet, read emails, summarize documents all in a very 06:16.650 --> 06:18.090 cohesive workflow. 06:18.730 --> 06:23.250 Each additional tool requires careful integration and coordination. 06:23.610 --> 06:25.730 There are also scaling problems. 06:26.330 --> 06:33.690 Current approaches require custom integration for each tool, creating maintenance burden and inconsistent 06:33.690 --> 06:35.130 user experience. 06:35.530 --> 06:39.130 This is how life has evolved over time. 06:39.570 --> 06:46.610 However, there is more room to evolve and that is where MCP can help us with some complexity and scaling 06:46.610 --> 06:47.490 problems. 06:48.170 --> 06:52.610 So let's take a deep dive into some other concepts before we learn about MCP. 06:52.970 --> 06:56.050 Thank you and I'll see you in the next video.