1 00:00:00,150 --> 00:00:03,180 -: As I mentioned at the beginning of this section, 2 00:00:03,180 --> 00:00:05,760 whether you know how it works or not, 3 00:00:05,760 --> 00:00:09,060 you can still get quality prompts 4 00:00:09,060 --> 00:00:12,330 and quality outputs from ChatGPT, 5 00:00:12,330 --> 00:00:16,590 but understanding these nuances and these subtleties 6 00:00:16,590 --> 00:00:20,130 is just gonna give you a more well-rounded education 7 00:00:20,130 --> 00:00:24,090 on how to effectively use it for your own personal life. 8 00:00:24,090 --> 00:00:27,090 So let's now talk about 9 00:00:27,090 --> 00:00:30,600 how does Chad GBT actually answer your questions. 10 00:00:30,600 --> 00:00:35,600 So GPT stands for Generative Pre-Trained Transformer, 11 00:00:37,020 --> 00:00:40,590 and it doesn't read words in the way that humans do, 12 00:00:40,590 --> 00:00:42,900 but it processes them 13 00:00:42,900 --> 00:00:47,100 through a series of computational steps. 14 00:00:47,100 --> 00:00:50,490 And this is gonna be an overview of how it computes 15 00:00:50,490 --> 00:00:54,360 and handles these words from your prompts. 16 00:00:54,360 --> 00:00:57,750 First, we have tokenization. 17 00:00:57,750 --> 00:01:01,680 Initially, the text that is inputted from a user 18 00:01:01,680 --> 00:01:04,950 is broken down into smaller pieces, 19 00:01:04,950 --> 00:01:09,450 much like splitting a sentence into individual words 20 00:01:09,450 --> 00:01:11,670 or even smaller parts. 21 00:01:11,670 --> 00:01:14,730 This process is called tokenization. 22 00:01:14,730 --> 00:01:19,260 For example, the sentence, Chatgpt is great, 23 00:01:19,260 --> 00:01:22,440 might be split down into three pieces, 24 00:01:22,440 --> 00:01:26,190 ChatGPT, is, and great. 25 00:01:26,190 --> 00:01:29,940 Each of these pieces is referred to as a token, 26 00:01:29,940 --> 00:01:33,540 and this token breakdown makes it easier for the model 27 00:01:33,540 --> 00:01:37,890 to analyze the text and understand its structure. 28 00:01:37,890 --> 00:01:41,160 In short, it's taking a large amount of information 29 00:01:41,160 --> 00:01:43,530 and breaking it down into smaller chunks 30 00:01:43,530 --> 00:01:46,710 that it can dissect and further understand 31 00:01:46,710 --> 00:01:48,990 by looking at things individually 32 00:01:48,990 --> 00:01:51,960 and then looking at how these different parts and pieces 33 00:01:51,960 --> 00:01:53,670 relate to each other. 34 00:01:53,670 --> 00:01:56,760 This leads to the embedding phase. 35 00:01:56,760 --> 00:02:00,150 After the text is broken down into smaller pieces, 36 00:02:00,150 --> 00:02:04,560 each piece is then converted into a numerical form, 37 00:02:04,560 --> 00:02:06,540 like a unique code. 38 00:02:06,540 --> 00:02:10,229 So this way, the computer can understand 39 00:02:10,229 --> 00:02:12,810 and work with it more efficiently. 40 00:02:12,810 --> 00:02:15,330 This process is known as embedding. 41 00:02:15,330 --> 00:02:17,670 And it's like translating the words 42 00:02:17,670 --> 00:02:20,070 that you inputted into a language 43 00:02:20,070 --> 00:02:21,690 that the model can understand, 44 00:02:21,690 --> 00:02:24,840 where each word or individual piece 45 00:02:24,840 --> 00:02:28,770 gets its own unique numerical identifier. 46 00:02:28,770 --> 00:02:31,710 This way, GPT can process the text 47 00:02:31,710 --> 00:02:35,880 in a format that's suitable for mathematical operations, 48 00:02:35,880 --> 00:02:38,190 which are used in the later steps 49 00:02:38,190 --> 00:02:42,270 to analyze the text and generate a response. 50 00:02:42,270 --> 00:02:44,160 This is necessary because these models 51 00:02:44,160 --> 00:02:48,810 do not understand words in the same way a human mind does. 52 00:02:48,810 --> 00:02:51,120 So with these unique codes, 53 00:02:51,120 --> 00:02:55,470 the numerical values are pertaining to semantic 54 00:02:55,470 --> 00:02:58,500 and syntactic information. 55 00:02:58,500 --> 00:03:03,150 Semantic information relates to the meaning of words 56 00:03:03,150 --> 00:03:06,420 and their relationship with other words. 57 00:03:06,420 --> 00:03:09,390 For instance, consider the words doctor, 58 00:03:09,390 --> 00:03:11,940 nurse, and hospital. 59 00:03:11,940 --> 00:03:14,520 Semantically, these words are related 60 00:03:14,520 --> 00:03:17,820 because they all pertain to healthcare. 61 00:03:17,820 --> 00:03:19,440 In the embedding space, 62 00:03:19,440 --> 00:03:22,470 they might be represented by different vectors 63 00:03:22,470 --> 00:03:24,360 that are close to each other, 64 00:03:24,360 --> 00:03:27,360 indicating their related meanings to each other. 65 00:03:27,360 --> 00:03:30,120 On the other hand, a word like pineapple, 66 00:03:30,120 --> 00:03:32,340 which is unrelated to healthcare, 67 00:03:32,340 --> 00:03:37,200 might have a vector far away from the words doctor, nurse, 68 00:03:37,200 --> 00:03:38,460 and hospital. 69 00:03:38,460 --> 00:03:42,720 So it's a Way of grouping information and relationships 70 00:03:42,720 --> 00:03:44,850 that words have with each other. 71 00:03:44,850 --> 00:03:46,892 For syntactic information, 72 00:03:46,892 --> 00:03:51,510 this is about the arrangements of the words and phrases 73 00:03:51,510 --> 00:03:55,470 to create formed sentences in a language. 74 00:03:55,470 --> 00:03:58,530 It's like the framework or rules that dictate 75 00:03:58,530 --> 00:04:01,470 how sentences are constructed. 76 00:04:01,470 --> 00:04:05,400 In English, the typical word ordering for a sentence 77 00:04:05,400 --> 00:04:10,400 is Subject-Verb-Object, or SVO. 78 00:04:10,410 --> 00:04:14,190 So in the sentence, John eats apples, 79 00:04:14,190 --> 00:04:18,089 John is the subject, eats is the verb, 80 00:04:18,089 --> 00:04:20,670 and apples is the object. 81 00:04:20,670 --> 00:04:24,510 Syntactic information helps identify 82 00:04:24,510 --> 00:04:29,510 the correct order of words that make meaningful sentences. 83 00:04:29,730 --> 00:04:32,910 In combination, these two types of information 84 00:04:32,910 --> 00:04:37,910 help GPT to understand the context and nuances 85 00:04:38,520 --> 00:04:40,680 of a user's input, 86 00:04:40,680 --> 00:04:44,910 which leads to step three, which is the encoding process. 87 00:04:44,910 --> 00:04:48,330 This is where tokens are fed into the GPT model 88 00:04:48,330 --> 00:04:52,980 and are essentially moving through a multi-layered network 89 00:04:52,980 --> 00:04:57,120 that refines the understanding of each token 90 00:04:57,120 --> 00:04:59,970 based on the tokens around it. 91 00:04:59,970 --> 00:05:02,880 So let's break down what this actually means. 92 00:05:02,880 --> 00:05:04,860 So first, we have our layers, 93 00:05:04,860 --> 00:05:07,980 and GPT has a series of these layers 94 00:05:07,980 --> 00:05:11,220 that are similar to the floors of a building, 95 00:05:11,220 --> 00:05:15,090 and each layer performs a special operation 96 00:05:15,090 --> 00:05:17,220 on the incoming tokens. 97 00:05:17,220 --> 00:05:20,814 So imagine each layer as a kind of workshop 98 00:05:20,814 --> 00:05:24,480 where tokens get refined, reshaped, 99 00:05:24,480 --> 00:05:29,100 based on the surrounding context of information. 100 00:05:29,100 --> 00:05:34,100 GPT-3, for example, has 175 billion parameters 101 00:05:35,190 --> 00:05:40,190 distributed across 93 transformer layers. 102 00:05:41,010 --> 00:05:45,150 Now, in each of these layers, tokens are transformed 103 00:05:45,150 --> 00:05:48,900 through very complex mathematical operations, 104 00:05:48,900 --> 00:05:52,500 and you can think of this as kind of a translation 105 00:05:52,500 --> 00:05:54,600 from one language to another, 106 00:05:54,600 --> 00:05:59,600 helping to build a richer understanding of each token. 107 00:06:00,090 --> 00:06:03,120 We then have the attention mechanism. 108 00:06:03,120 --> 00:06:07,680 So within each layer, there's a mechanism called attention, 109 00:06:07,680 --> 00:06:09,780 which allows each token 110 00:06:09,780 --> 00:06:14,780 to kind of look at other tokens in the input 111 00:06:15,210 --> 00:06:18,360 and adjust its own representation 112 00:06:18,360 --> 00:06:21,660 based on what it, quote, unquote, "sees," 113 00:06:21,660 --> 00:06:23,880 from looking at other tokens. 114 00:06:23,880 --> 00:06:28,880 For instance, the word bank might adjust its representation 115 00:06:29,160 --> 00:06:32,250 based on whether the surrounding words 116 00:06:32,250 --> 00:06:36,300 relate to a financial institution bank 117 00:06:36,300 --> 00:06:39,570 or the side of a river bank. 118 00:06:39,570 --> 00:06:41,790 This type of attention mechanism 119 00:06:41,790 --> 00:06:44,461 is incredibly important for GPT 120 00:06:44,461 --> 00:06:49,461 to understand how all of the inputted words 121 00:06:49,950 --> 00:06:51,360 relate to each other. 122 00:06:51,360 --> 00:06:56,040 This is all leading to what's called contextual adjustment. 123 00:06:56,040 --> 00:06:58,770 Now, the attention mechanism 124 00:06:58,770 --> 00:07:03,270 is helping in adjusting the representation 125 00:07:03,270 --> 00:07:06,540 of each token based on its context. 126 00:07:06,540 --> 00:07:11,540 So this is to make sure that the model understands each word 127 00:07:11,760 --> 00:07:13,500 in a way that makes sense, 128 00:07:13,500 --> 00:07:17,070 given all of the surrounding words around it. 129 00:07:17,070 --> 00:07:20,970 So now these more refined tokens 130 00:07:20,970 --> 00:07:23,460 move up to the next layer, 131 00:07:23,460 --> 00:07:26,130 and they go through a similar process 132 00:07:26,130 --> 00:07:27,900 again and again and again, 133 00:07:27,900 --> 00:07:32,220 with each of these layers that it passes through, 134 00:07:32,220 --> 00:07:35,490 there's a deeper understanding of the context 135 00:07:35,490 --> 00:07:38,463 and the information that the model gains 136 00:07:38,463 --> 00:07:40,500 through this process. 137 00:07:40,500 --> 00:07:43,890 So as these tokens are moving through these layers, 138 00:07:43,890 --> 00:07:48,150 they aggregate and gather more and more information 139 00:07:48,150 --> 00:07:50,040 surrounding the original tokens, 140 00:07:50,040 --> 00:07:53,640 which helps to continuously build a rich 141 00:07:53,640 --> 00:07:55,260 and contextual understanding 142 00:07:55,260 --> 00:07:58,410 of the entire output as a whole. 143 00:07:58,410 --> 00:08:01,800 So by the time the tokens reach the final layer, 144 00:08:01,800 --> 00:08:04,770 they have been significantly refined 145 00:08:04,770 --> 00:08:09,150 and carry a detailed understanding of the input text 146 00:08:09,150 --> 00:08:12,360 provided by the surrounding words. 147 00:08:12,360 --> 00:08:14,250 Each layer has contributed 148 00:08:14,250 --> 00:08:17,357 to building this deep understanding of the input, 149 00:08:17,357 --> 00:08:20,700 making the tokens ready for the next step, 150 00:08:20,700 --> 00:08:23,160 which is the output generation. 151 00:08:23,160 --> 00:08:25,470 After this multi-layered process 152 00:08:25,470 --> 00:08:28,440 of processing the user's input, 153 00:08:28,440 --> 00:08:33,440 GPT begins the task of creating a response. 154 00:08:33,480 --> 00:08:36,990 It starts with the tokens provided in the input 155 00:08:36,990 --> 00:08:41,990 and then works to predict what should come next 156 00:08:42,299 --> 00:08:44,520 one token at a time. 157 00:08:44,520 --> 00:08:47,100 So let's dive into this process. 158 00:08:47,100 --> 00:08:51,390 So we start with our prediction and selection. 159 00:08:51,390 --> 00:08:52,950 For each new token, 160 00:08:52,950 --> 00:08:56,850 GPT looks at all the tokens that have come before it, 161 00:08:56,850 --> 00:08:59,220 which includes the user's input 162 00:08:59,220 --> 00:09:01,710 on top of any other tokens 163 00:09:01,710 --> 00:09:05,820 that were previously generated inside that specific chat, 164 00:09:05,820 --> 00:09:10,410 and this is gonna help predict the most likely next token. 165 00:09:10,410 --> 00:09:13,350 It makes this prediction based on patterns 166 00:09:13,350 --> 00:09:16,080 it has learned from the vast amount of data 167 00:09:16,080 --> 00:09:17,670 it was trained on, 168 00:09:17,670 --> 00:09:21,720 and then GPT selects the one that it calculates 169 00:09:21,720 --> 00:09:24,900 to be the most likely next token 170 00:09:24,900 --> 00:09:28,260 based on its understanding from the training data. 171 00:09:28,260 --> 00:09:31,650 So when processing the input phrase, 172 00:09:31,650 --> 00:09:35,383 Albert Einstein was the world's most..., 173 00:09:36,299 --> 00:09:38,640 GPT would go through the following steps 174 00:09:38,640 --> 00:09:41,790 to predict and select the next word. 175 00:09:41,790 --> 00:09:44,160 First, it would go through the tokenization process 176 00:09:44,160 --> 00:09:45,180 that we mentioned. 177 00:09:45,180 --> 00:09:48,540 The input phrase is tokenized into individual tokens, 178 00:09:48,540 --> 00:09:52,170 aka Albert is separate, Einstein is separate, 179 00:09:52,170 --> 00:09:57,170 was, the, world's, and most are all individually looked at. 180 00:09:57,480 --> 00:09:58,770 Then we have embedding, 181 00:09:58,770 --> 00:10:02,370 where each token is converted into a numerical value. 182 00:10:02,370 --> 00:10:05,940 Those values are then processed through the multiple layers 183 00:10:05,940 --> 00:10:07,530 of the GPT model. 184 00:10:07,530 --> 00:10:09,900 Through the self-attention mechanism, 185 00:10:09,900 --> 00:10:12,510 the model then identifies the relationship 186 00:10:12,510 --> 00:10:14,040 between those tokens, 187 00:10:14,040 --> 00:10:17,220 understanding, for instance, that Albert Einstein 188 00:10:17,220 --> 00:10:21,510 is referring to a notable individual in history. 189 00:10:21,510 --> 00:10:24,570 We then have the probability distribution. 190 00:10:24,570 --> 00:10:26,400 Based on the context, 191 00:10:26,400 --> 00:10:29,760 the model computes a probability distribution 192 00:10:29,760 --> 00:10:34,590 over the vocabulary for the next token 193 00:10:34,590 --> 00:10:36,540 after the word most. 194 00:10:36,540 --> 00:10:39,900 Some common continuations for this probability 195 00:10:39,900 --> 00:10:43,710 would likely be most famous, most influential, 196 00:10:43,710 --> 00:10:45,000 most brilliant. 197 00:10:45,000 --> 00:10:48,840 These would all receive high probabilities. 198 00:10:48,840 --> 00:10:52,020 Now, what exactly is a probability distribution? 199 00:10:52,020 --> 00:10:54,510 It's simply put is like a rule 200 00:10:54,510 --> 00:10:58,530 that tells you how likely different outcomes are 201 00:10:58,530 --> 00:11:03,530 in a situation involving some sort of chance or randomness. 202 00:11:03,900 --> 00:11:07,620 So imagine you have a jar of colored candies, 203 00:11:07,620 --> 00:11:10,170 red, blue, and green. 204 00:11:10,170 --> 00:11:13,830 If the candies are mixed together all completely equally, 205 00:11:13,830 --> 00:11:16,110 the chance of you pulling out one color 206 00:11:16,110 --> 00:11:18,840 is gonna be the same, which is one out of three, 207 00:11:18,840 --> 00:11:23,220 or 33.33333333%. 208 00:11:23,220 --> 00:11:26,640 Now, if you write down these chances for each color, 209 00:11:26,640 --> 00:11:29,760 we get a simple probability distribution. 210 00:11:29,760 --> 00:11:34,760 The chance of red is 33.3, the chance of blue is 33.3, 211 00:11:35,160 --> 00:11:38,538 and the chance of green is 33.3. 212 00:11:38,538 --> 00:11:42,450 In the case of GPT, generating a word, 213 00:11:42,450 --> 00:11:46,200 think of it like the model has a huge jar of words, 214 00:11:46,200 --> 00:11:49,380 and some words are more likely to be picked 215 00:11:49,380 --> 00:11:52,170 based on the previous words. 216 00:11:52,170 --> 00:11:56,010 The probability distribution is the rule that tells GPT 217 00:11:56,010 --> 00:11:59,040 how likely each word is to be picked next, 218 00:11:59,040 --> 00:12:01,980 Simply put, the probability distribution 219 00:12:01,980 --> 00:12:04,320 is the rule that tells GPT 220 00:12:04,320 --> 00:12:07,980 how likely each word is to be picked next. 221 00:12:07,980 --> 00:12:12,390 So GPT goes through this process, it chooses the next word, 222 00:12:12,390 --> 00:12:14,100 and then we have continuation, 223 00:12:14,100 --> 00:12:16,890 where the model continues this process 224 00:12:16,890 --> 00:12:21,630 for each subsequent token using the growing context 225 00:12:21,630 --> 00:12:25,590 to inform the prediction and selection process 226 00:12:25,590 --> 00:12:26,880 of the next token. 227 00:12:26,880 --> 00:12:31,170 For instance, after choosing the word influential, 228 00:12:31,170 --> 00:12:34,020 it might predict next that the proper word 229 00:12:34,020 --> 00:12:37,170 would be physicist, with a high probability, 230 00:12:37,170 --> 00:12:39,720 given that the context of the prompt 231 00:12:39,720 --> 00:12:42,030 is about Albert Einstein. 232 00:12:42,030 --> 00:12:44,820 With this, the model continues generating tokens 233 00:12:44,820 --> 00:12:47,640 until a stopping criteria is met, 234 00:12:47,640 --> 00:12:50,700 such as the end of a sentence, punctuation, 235 00:12:50,700 --> 00:12:54,090 or the maximum token limit has been reached. 236 00:12:54,090 --> 00:12:55,830 This process continues 237 00:12:55,830 --> 00:12:59,010 until the model would likely create the phrase, 238 00:12:59,010 --> 00:13:03,150 Albert Einstein was the world's most influential physicist 239 00:13:03,150 --> 00:13:04,770 of the 20th century, 240 00:13:04,770 --> 00:13:07,650 whose work revolutionize our understanding 241 00:13:07,650 --> 00:13:10,800 of the fundamental laws of the universe. 242 00:13:10,800 --> 00:13:12,720 So it's really incredible to know 243 00:13:12,720 --> 00:13:16,290 that when you put in a prompt and instantly get your output, 244 00:13:16,290 --> 00:13:19,560 this is the process that's going on underneath the hood 245 00:13:19,560 --> 00:13:22,290 that involves billions of parameters, 246 00:13:22,290 --> 00:13:24,570 training data and information, 247 00:13:24,570 --> 00:13:27,990 that is synthesizing all this complex information 248 00:13:27,990 --> 00:13:30,453 and giving you your desired output.