WEBVTT 00:00.184 --> 00:02.790 Instructor: First, it's important to start with 00:02.790 --> 00:04.620 what is language modeling? 00:04.620 --> 00:07.893 So this is the formal definition from Wikipedia, 00:08.820 --> 00:11.970 which talks about a distribution of probabilities 00:11.970 --> 00:13.890 over a sequence of words. 00:13.890 --> 00:17.340 So if I'm going to bring this down to earth, 00:17.340 --> 00:20.910 language modeling is the task of predicting 00:20.910 --> 00:23.460 what word will come up next. 00:23.460 --> 00:24.750 You can think about it 00:24.750 --> 00:28.140 like a super, super smart autocomplete. 00:28.140 --> 00:31.800 So if we have a sentence like the dog wigged its, 00:31.800 --> 00:34.230 and then a collection of words, 00:34.230 --> 00:36.300 then according to this sentence, 00:36.300 --> 00:39.960 the language model for each word that you see right now 00:39.960 --> 00:41.640 has a probability. 00:41.640 --> 00:43.800 And what it would output for you 00:43.800 --> 00:47.174 is the highest probability, that it's correct. 00:47.174 --> 00:50.340 So now let's review the formal definition 00:50.340 --> 00:52.590 of a language model. 00:52.590 --> 00:56.070 So for that, you don't really need to know computers 00:56.070 --> 00:59.103 and science, you'll see it's very, very intuitive. 00:59.940 --> 01:03.330 So given a sequence of words, 01:03.330 --> 01:07.260 we'll mark the words X1, X2, and Xt. 01:07.260 --> 01:09.990 So you can simply map each X 01:09.990 --> 01:12.990 to the word in the sentence previously. 01:12.990 --> 01:15.330 We want to compute the probability distribution 01:15.330 --> 01:18.270 of the next word, which is Xt+1. 01:18.270 --> 01:21.720 Because we had so far X1 until Xt, 01:21.720 --> 01:23.640 those are the words in our sentence, 01:23.640 --> 01:26.040 and we want to know what is the probability 01:26.040 --> 01:28.950 of the word Xt+1. 01:28.950 --> 01:29.783 Cool. 01:29.783 --> 01:32.493 So P is the notation for probability. 01:33.360 --> 01:35.977 So in this notation, you see right now we want to ask, 01:35.977 --> 01:40.230 "Hey, what is the probability of the next word, 01:40.230 --> 01:43.980 which is Xt+1, given this is the long line 01:43.980 --> 01:45.060 that you see right here, 01:45.060 --> 01:49.950 that we had the sentence X1 until Xt before. 01:49.950 --> 01:54.180 And Xt+1 needs to be a word in our vocabulary." 01:54.180 --> 01:56.790 Now, our vocabulary is an object, 01:56.790 --> 02:00.450 which is notated as V for vocabulary. 02:00.450 --> 02:04.110 So if we just summarize what is a language model, 02:04.110 --> 02:08.431 then it's simply the idea that we have a sentence, 02:08.431 --> 02:11.760 a couple of words that come together, one after another, 02:11.760 --> 02:14.790 and we want to guess what will be the next word. 02:14.790 --> 02:17.883 So it's super, super similar like autocomplete, 02:18.780 --> 02:21.450 just like we know from our day-to-day lives, 02:21.450 --> 02:25.440 like when we write a text message and we have a suggestion, 02:25.440 --> 02:27.630 so it's based on a language model. 02:27.630 --> 02:30.137 The same thing when we search for something 02:30.137 --> 02:31.743 in our search engine. 02:33.330 --> 02:35.880 So what's a large language model 02:35.880 --> 02:37.950 which is all the hype right now? 02:37.950 --> 02:42.750 So a large language model or abbreviated, LLM, 02:42.750 --> 02:45.870 is simply a language model like we saw before 02:45.870 --> 02:50.870 that was trained on a huge amount of data. 02:52.740 --> 02:55.620 So you can think about it as a model, 02:55.620 --> 02:58.920 a language model that was trained on so much data 02:58.920 --> 03:01.980 that is pretty much very, very good 03:01.980 --> 03:04.383 at calculating those probabilities. 03:06.060 --> 03:08.250 So every time you write a prompt, 03:08.250 --> 03:11.550 then what you basically do is give an input of words 03:11.550 --> 03:16.410 for the LLM and the LLM simply now guesses 03:16.410 --> 03:19.980 and make its best effort in order to output you 03:19.980 --> 03:22.140 the next word, one after another. 03:22.140 --> 03:23.820 So it calculates the probability 03:23.820 --> 03:26.700 and gives you the word with the highest probability 03:26.700 --> 03:29.400 in the context of the input you provided. 03:29.400 --> 03:33.480 So it simply guesses word after word after word, 03:33.480 --> 03:34.950 what you are expecting. 03:34.950 --> 03:38.812 Now, this is why LLMs can sometimes output us something 03:38.812 --> 03:43.110 that is so farfetched from reality and simply not true, 03:43.110 --> 03:46.020 because it's simply guessing the probability 03:46.020 --> 03:48.480 and simply relying on that. 03:48.480 --> 03:51.063 So that's the entire concept of LLM.