WEBVTT 00:00.240 --> 00:05.240 In this video, we'll cover the most important information about large language models. 00:05.240 --> 00:09.520 And this knowledge will be very essential because we'll talk about how they are built. 00:09.560 --> 00:12.120 What even is a large language model. 00:12.120 --> 00:16.760 We'll talk about the fine tuning, how large language models are creating the responses, and a lot 00:16.800 --> 00:17.400 of more. 00:17.640 --> 00:21.720 And now there is a question why do we want to learn about that? 00:21.760 --> 00:23.360 And actually let me explain. 00:23.400 --> 00:31.080 When we create our workflows automations inside here and then we are using agents which are purely LM 00:31.080 --> 00:31.680 models. 00:31.680 --> 00:36.800 So for example, we are using ChatGPT, deep seq and other different applications. 00:36.920 --> 00:40.840 And when you understand how they work, how actually it works. 00:40.840 --> 00:47.320 So LM models, you are able to create better systems, more sophisticated agents and in general better 00:47.320 --> 00:48.200 projects. 00:48.200 --> 00:53.880 So again this video can be quite long over here and will cover all of the information you need to know 00:54.040 --> 00:55.560 about LM models. 00:55.560 --> 00:57.560 And again it will be very important. 00:57.600 --> 00:59.720 So with all that being said let's dive in. 00:59.720 --> 00:59.810 Stephen. 01:00.010 --> 01:00.490 All right. 01:00.490 --> 01:02.650 So let's start with the most basic question. 01:02.650 --> 01:04.010 What is an LM. 01:04.130 --> 01:05.650 And it's a large language model. 01:05.650 --> 01:11.090 And here actually we can read LM is an advanced type of artificial intelligence that can understand 01:11.090 --> 01:16.570 and generate human like text alarms, power tools like ChatGPT, cloud Deep Seek and many others. 01:16.730 --> 01:22.050 So here for an example, we've got Grok and Google Gemini, ChatGPT and again cloud. 01:22.170 --> 01:24.130 So these are LM models. 01:24.130 --> 01:27.810 And again we are using them actually with our agents. 01:27.810 --> 01:29.610 So you know we are using these models. 01:29.610 --> 01:35.170 And here they are trained on huge amounts of text from the internet books and more. 01:35.290 --> 01:38.090 So I like to compare that to a big library. 01:38.090 --> 01:42.970 And not only one however, you know, thousands of them, or even sometimes more. 01:43.170 --> 01:47.530 So LM model is built from huge amount of the data. 01:47.730 --> 01:49.690 So it contains all of the data. 01:50.170 --> 01:51.290 It's trained on that. 01:51.290 --> 01:57.930 And actually it's very powerful because actually imagine inside LM you can have all of the information 01:58.490 --> 02:05.490 from a lot of big libraries, like any articles, any, you know, blog posts, like a lot of a lot 02:05.530 --> 02:06.170 of data. 02:06.210 --> 02:08.050 However about that also later. 02:08.450 --> 02:13.330 So they learned to answer questions, write stories, summarize translate code and much more. 02:13.650 --> 02:15.130 So again it's also powerful. 02:15.130 --> 02:21.370 You've got kind of the model that can think and also is trained actually on a huge amount of the data. 02:21.650 --> 02:28.850 And even, you know, like models like bit like people and thinking and reasoning processes and even 02:28.850 --> 02:33.850 for now, we had a lot of different, um, you know, research about the health care and diagnosing 02:33.850 --> 02:35.490 different problems and so on. 02:35.490 --> 02:40.450 So, I mean, like these like, models are so, so powerful. 02:40.810 --> 02:45.170 Um, however, you will see that like, um, you know, like later in this video. 02:45.490 --> 02:46.570 Um, let's go ahead. 02:47.290 --> 02:51.250 And here we've got another topic, however, LM builds. 02:51.290 --> 02:56.090 And I can say it's not that hard to comprehend because we've got only two files. 02:56.090 --> 02:59.740 So at its core, an LM is surprisingly simple to run. 02:59.780 --> 03:01.100 We've got only two files. 03:01.100 --> 03:02.980 The first one is parameters file. 03:03.340 --> 03:03.820 I'm actually. 03:03.820 --> 03:08.340 I will explain that more later after we cover like more topics. 03:08.340 --> 03:11.940 For example I'm how LM is creating the responses. 03:12.100 --> 03:15.740 However, this is the giant file containing everything the model knows. 03:15.740 --> 03:19.100 It's the result of training on huge data on huge datasets. 03:19.140 --> 03:23.980 For a model like llama M2 70 billion parameters. 03:23.980 --> 03:25.780 So this is 70 billion. 03:26.060 --> 03:29.620 This file can be over 140GB. 03:29.820 --> 03:31.420 So it's a huge file. 03:31.620 --> 03:35.060 And yeah like we've got a huge data sets stored. 03:35.060 --> 03:37.140 And the weight is very big. 03:37.540 --> 03:42.700 And the second file is run a runner file which is also the code file. 03:42.820 --> 03:49.060 So short script C, Python or other languages that leads the parameters and runs the model letting you 03:49.060 --> 03:50.060 interact with it. 03:50.660 --> 03:53.740 So actually you know we've got only two files. 03:53.740 --> 03:55.860 So there you've got the analogy. 03:56.060 --> 03:59.550 Think of of the parameters file as the brain. 03:59.550 --> 04:03.390 So it's the brain actually here memory and experience. 04:03.710 --> 04:05.910 And then the runner file as the body. 04:05.910 --> 04:08.230 So code to make it work. 04:08.470 --> 04:12.950 So actually the runner file actually takes you know all of the knowledge here, all of the information 04:12.950 --> 04:13.670 from there. 04:13.830 --> 04:16.950 And like does the specific actions you know. 04:17.150 --> 04:20.950 So this is very simple I think it's not that hard to comprehend. 04:20.950 --> 04:28.670 So we've got actually our database huge database data sets here inside the parameters file. 04:28.670 --> 04:34.630 And actually here the runner file which is doing some specific tasks based on the parameters file. 04:34.630 --> 04:36.630 All right let's go ahead. 04:36.830 --> 04:38.870 And now there is a very important topic. 04:38.910 --> 04:40.790 How does an LLM think. 04:41.030 --> 04:48.590 Because it's very tricky sometimes actually we can think that LLM is actually thinking as the real human. 04:48.630 --> 04:55.150 However this is not the truth because LLM can't understand what you are typing and instead they are 04:55.150 --> 04:59.600 basing on the weights on actually different, you know, things. 04:59.720 --> 05:03.400 However, let me actually explain that lambs don't think like humans. 05:03.400 --> 05:09.040 They predict the next word in a sentence, one word at a time, using everything they learned during 05:09.080 --> 05:09.760 training. 05:09.840 --> 05:13.440 So example the Eiffel Tower is located in. 05:13.440 --> 05:18.880 So here the LM analyzes the context and assigns probabilities to possible next words. 05:19.040 --> 05:21.560 So it can't understand this text however. 05:21.800 --> 05:25.840 And we've got a lot of different parameters huge data sets. 05:25.880 --> 05:28.000 And now we actually actually imagine this. 05:28.000 --> 05:30.720 By this way the Eiffel Tower is located in. 05:30.760 --> 05:35.480 So the LM analyzes the context and assigns probabilities to possible next words. 05:35.480 --> 05:37.520 And now imagine the following situation. 05:37.560 --> 05:39.840 You've seen, let's say ten articles. 05:39.840 --> 05:40.280 All right. 05:40.280 --> 05:43.480 In your career you've seen ten articles like whatever. 05:43.640 --> 05:51.160 And here every time whenever you had the word actually like this sentence, um, like you've seen the 05:51.160 --> 05:52.240 word Paris next. 05:52.240 --> 05:52.880 All right. 05:52.880 --> 05:55.840 So llms work and by this way. 05:55.840 --> 06:02.840 So they predict what is actually I can say what fits to the to the previous word and to the entire sentence, 06:02.840 --> 06:03.400 you know. 06:03.400 --> 06:10.080 So actually here, um, we've got the Eiffel Tower is located in and actually on the database and in 06:10.080 --> 06:10.960 the data sets. 06:11.160 --> 06:13.400 And mostly we had the word which is Paris. 06:13.400 --> 06:14.120 All right. 06:14.160 --> 06:18.960 And here, um, the probability of that is 97%. 06:19.280 --> 06:25.520 And the next word is France, the second is the third is London, and the fourth is New York. 06:25.840 --> 06:27.360 Um, so we've got the explanation. 06:27.360 --> 06:33.200 The model will most likely choose Paris as the next word because it learned this, um, is the best 06:33.200 --> 06:33.760 fit. 06:33.760 --> 06:37.440 So from huge amount of databases, data sets and more. 06:38.080 --> 06:43.800 Um, however, there we've got some very important actually a metric which is called temperature. 06:43.960 --> 06:50.360 So sometimes when we have like when we have our models, we can actually adjust the temperature. 06:50.520 --> 06:53.120 And it's I can say it's the creativeness. 06:53.120 --> 06:59.450 So whenever the temperature is higher, like the model is more creative and might surprise you with 06:59.490 --> 07:01.130 unusual words or phrases. 07:01.250 --> 07:06.730 So imagine that us like Paris, suddenly and has less percent. 07:06.730 --> 07:10.010 And for an example, France has like more percent, you know. 07:10.250 --> 07:16.490 So here when you have higher temperature inside the model, it produces just more creative responses, 07:16.490 --> 07:17.970 just more unusual. 07:18.010 --> 07:20.890 However, with the low temperature, it's very predictable. 07:20.890 --> 07:26.570 So the model is more conservative, always picks the most likely next word and the example. 07:26.570 --> 07:30.010 So with low temperature the cat sat on the mat. 07:30.010 --> 07:33.330 And here with high temperature the cat sat on the piano. 07:34.010 --> 07:35.330 So it works by this way. 07:35.370 --> 07:40.810 However, the most important fact is that, um, LM doesn't think like the human. 07:40.970 --> 07:46.090 However, it takes actually different weights, um, from the from the data they learned. 07:46.410 --> 07:48.410 And here, um, yeah. 07:48.610 --> 07:51.410 And here actually it decides what should be next. 07:51.410 --> 07:53.210 So again you've got the probabilities. 07:53.460 --> 07:58.700 And now let's go back to our parameters file because it's very important. 07:58.940 --> 08:04.620 And every time the model makes a prediction actually it checks if it was correct or right. 08:04.740 --> 08:11.620 And if it wasn't it slightly like tweaks um, like the parameter um, which is a weight. 08:11.820 --> 08:14.820 Um, so it gets a little better next time. 08:15.060 --> 08:17.100 So parameters are weights. 08:17.100 --> 08:18.420 This is not the text however. 08:18.460 --> 08:19.300 These are weights. 08:19.540 --> 08:24.780 And every time whenever we have like, you know, like the answer it tweaks. 08:24.940 --> 08:25.500 Um, yeah. 08:25.660 --> 08:27.580 It tweaks em to the perfection. 08:27.700 --> 08:33.380 So like the fact from that is, is that like LM models are still being improved? 08:33.460 --> 08:37.060 Um, however about that later in the fine tuning and improving the model. 08:37.220 --> 08:43.460 So let's go ahead right here to the fourth point which is training LMS. 08:43.700 --> 08:45.820 So it's like compressing the internet. 08:46.220 --> 08:50.460 Um, actually I gave you a comparison to a library. 08:50.500 --> 08:54.140 However, internet is also a great comparison. 08:54.140 --> 08:56.500 So on the internet you've got a lot of data. 08:56.500 --> 09:02.460 And here training an LLM is like compressing a huge chunk of the internet into a brain of parameters. 09:02.580 --> 09:07.660 Companies like meta, OpenAI, Google and deep sea train models using massive data, data, datasets, 09:07.700 --> 09:12.020 website books, articles, sometimes ten terabytes or more. 09:12.940 --> 09:14.580 So it's so, so great. 09:15.020 --> 09:18.060 Um, I've got for you some, um, nice actually. 09:18.060 --> 09:21.060 Note because Deep SIG did something differently. 09:21.340 --> 09:24.460 Um, however, I, I explained that in my deep SIG masterclass. 09:24.460 --> 09:26.660 Nevertheless, um, in a while about that. 09:26.860 --> 09:33.300 Um, so training requires thousands of GPUs for days or weeks, costing millions of dollars. 09:33.460 --> 09:36.140 So the cost is really, really high. 09:36.500 --> 09:38.900 The result is a huge parameters file. 09:38.900 --> 09:42.220 It's not a copy of the internet, but a lossy compression. 09:42.260 --> 09:45.100 It remembers patterns, facts and styles. 09:45.100 --> 09:46.220 And here is the note. 09:46.260 --> 09:51.550 Because actually, you know, common llms are trained on the data from internet. 09:51.590 --> 09:54.670 It takes a huge actually GPU. 09:55.110 --> 09:59.830 And also, you know, the space like the cost is very, very high. 10:00.150 --> 10:07.670 However, Deep Seek was released and it was a total surprise for everyone because what they did, they 10:07.710 --> 10:13.550 actually trained their models not on the data from the internet, articles, books, like huge data 10:13.590 --> 10:14.030 sets. 10:14.350 --> 10:18.670 However, they trained its model on other models. 10:18.670 --> 10:26.150 So actually they took, um, they work in simple words and they trained um, its model on that. 10:26.150 --> 10:33.110 So actually the cost was, um, you know, actually $6 million, which is just, um, a funny amount. 10:33.150 --> 10:33.510 All right. 10:33.550 --> 10:35.110 For training models. 10:35.390 --> 10:36.710 Um, however, let's read it. 10:36.710 --> 10:41.790 So models like Deep Six unique training methods or data sources, but the core concept remains trained 10:41.830 --> 10:45.230 on huge data to create a general purpose language engine. 10:45.950 --> 10:48.150 So again, it's very important. 10:48.350 --> 10:55.360 And I'm like training LLM LM is taking all of the data, information, putting in insight and using 10:55.360 --> 10:58.360 that, you know, actually to work do some specific tasks. 10:58.960 --> 11:01.960 Next what we have open source versus closed models. 11:01.960 --> 11:06.120 So open source models weights and code are available for everyone. 11:06.120 --> 11:10.440 You can download, read, modify or even fine tune them for custom uses. 11:10.920 --> 11:16.960 And what's very important, even when you have actually an open source model, you can download that 11:16.960 --> 11:18.080 on your computer. 11:18.200 --> 11:22.280 And that means you can even run it without internet connection. 11:22.320 --> 11:26.120 It runs on your actual computer locally. 11:26.280 --> 11:28.360 And here we've got you've got examples. 11:28.360 --> 11:32.560 So llama two, Mistral, Deep Sea Zephyr V3 Microsoft. 11:33.280 --> 11:41.400 And what is even very, very important actually when you run your model on your computer actually you 11:41.440 --> 11:42.960 keep all of the information. 11:43.160 --> 11:43.720 Um, yeah. 11:43.760 --> 11:44.480 Private. 11:44.680 --> 11:51.920 However, when you have closed models you're working on, for example, on OpenAI servers and here anthropic 11:51.920 --> 11:53.240 or even deep sea servers. 11:53.240 --> 11:56.760 And all of the data you provide is sent over to these companies. 11:56.760 --> 11:59.600 So let's say to China, United States and so on. 11:59.800 --> 12:06.480 So, um, actually, when you want to be very safe with what you type inside the model, um, you can 12:06.520 --> 12:08.280 like download Deep Seek on your computer. 12:08.320 --> 12:08.880 All right. 12:08.880 --> 12:09.760 You can do this. 12:09.760 --> 12:15.000 And then you can be 100% sure that data won't be sent anywhere. 12:15.680 --> 12:18.800 However, now let's talk about what are the close models. 12:18.800 --> 12:23.520 So you can you can access them only through a web interface or API. 12:23.880 --> 12:30.480 Code and weights are not public used via subscription or licensing examples ChatGPT cloud Gemini. 12:30.600 --> 12:35.640 So the most popular models however um, like it's very important to know the difference. 12:36.240 --> 12:41.040 Now what we have fine tuning how Llms are improved. 12:41.040 --> 12:48.360 So after initial training companies and now users can fine tune Llms to become better assistants, specialists 12:48.360 --> 12:49.450 or agents. 12:49.610 --> 12:50.530 Fine tuning. 12:50.530 --> 12:53.210 Start with a base LM general knowledge. 12:53.250 --> 12:59.690 Train further on custom datasets like Q and R pairs conversations or company documents. 12:59.730 --> 13:04.610 Also, you know models are trained on users responses. 13:04.650 --> 13:04.970 Yeah. 13:05.010 --> 13:11.530 So if you type the response, you provide the data and actually the parameters are improved and it's 13:11.530 --> 13:13.410 getting better over time. 13:13.570 --> 13:19.450 So the model becomes specialized better at helping following instructions or doing specific tasks. 13:19.570 --> 13:21.410 So imagine that as the experience. 13:21.450 --> 13:21.770 Yeah. 13:21.810 --> 13:22.850 As the experience. 13:22.850 --> 13:24.170 You work for an example. 13:24.170 --> 13:29.370 And the job for um, let's say one month or even ten years, you know. 13:29.410 --> 13:30.810 So this is the huge gap. 13:31.130 --> 13:32.650 Why is that important? 13:32.690 --> 13:38.330 It turns a generic document generator into a helpful assistant, customer support bot, coding helper 13:38.330 --> 13:40.250 or expert in a field. 13:40.410 --> 13:45.730 So it works by this way, it's very, very important to know how llms are improved. 13:45.730 --> 13:48.780 So they just get, over time more data. 13:48.780 --> 13:51.060 And by that they have like the feedback. 13:51.060 --> 13:51.660 All right. 13:51.660 --> 13:53.220 They collect the feedback. 13:53.220 --> 13:57.140 And on the feedback they get better and they produce better results. 13:57.820 --> 14:00.140 Yeah it's a similar situation. 14:00.140 --> 14:03.700 Whenever you know you've got the chat actually let me change the camera. 14:03.940 --> 14:08.580 Let's say you've got the chat with the ChatGPT and it produces for you about responses. 14:08.580 --> 14:11.860 And you say, all right, like give me something else. 14:11.860 --> 14:14.900 I don't like the, um, you know, uppercase. 14:14.900 --> 14:19.380 I don't like, let's say bullet points and then you like train the model. 14:19.420 --> 14:22.020 All right by but within your entire chat. 14:22.020 --> 14:26.940 So it's like a small comparison of how the models are trained. 14:26.980 --> 14:30.940 Of course, it's more advanced, however, you know, um, just a small digression. 14:30.980 --> 14:34.380 Llms as agents, real world applications. 14:34.660 --> 14:41.180 And now we go to the very important step, actually, because we are using llms inside Nar den. 14:41.340 --> 14:44.340 So when we create agents we use them. 14:44.460 --> 14:48.310 And now actually, you know, you have a big knowledge about them. 14:48.350 --> 14:48.750 All right. 14:48.790 --> 14:49.830 After this video. 14:50.230 --> 14:55.630 And I can say it's very important because Moodle LMS aren't just chatbots with tools like Zapier. 14:55.670 --> 14:59.350 And then another automation platforms, LMS become agents. 14:59.390 --> 15:06.070 They can connect to APIs, control apps, automate tasks, answer customer emails, summarize documents, 15:06.070 --> 15:06.910 and more. 15:07.070 --> 15:13.630 With Retrieval argument generation Rack, they can reference live documents and compound knowledge. 15:13.950 --> 15:20.710 LMS can be fine tuned or plugged into workflows, making them work for businesses, creators, developers, 15:20.710 --> 15:21.470 and more. 15:21.830 --> 15:25.870 So again, we are using LMS models such as ChatGPT. 15:26.550 --> 15:29.430 Here we are using Gemini. 15:29.590 --> 15:33.590 So again I'm kind of in the videos we are using LM models. 15:33.870 --> 15:41.110 And again it's very important to comprehend that that um yeah that it works by the way, not the other 15:41.270 --> 15:46.670 because otherwise you create just words, you know, words agents, words, Wars projects. 15:46.670 --> 15:52.990 And it's very important to again comprehend that because we based we actually based a lot of different 15:52.990 --> 15:55.030 workflows, automations on that. 15:55.750 --> 15:59.110 And last but not least, we've got the summary. 15:59.110 --> 16:06.550 So the power of LMS, LMS are advanced toward predictors powered by billions of parameters trained on 16:06.590 --> 16:09.390 huge data running from just two files. 16:09.590 --> 16:15.270 Open source LMS are democratizing AI, letting anyone experiment, run and improve models. 16:15.390 --> 16:21.390 Fine tuning and automation make LMS real world agents not just chatbots, but powerful AI tools for 16:21.390 --> 16:22.310 any task. 16:22.830 --> 16:29.830 So after this video, we have the great knowledge of what Lmx is, how they work and how they are built 16:29.870 --> 16:34.110 in general, and how they are improved and a lot more actually. 16:34.110 --> 16:35.230 I hope you enjoyed it. 16:35.230 --> 16:41.150 And now we can use this knowledge to create, again, better automations workflows inside. 16:41.190 --> 16:45.310 Nadine, thank you for watching for now and I will see you in the next video.