WEBVTT 1 00:00:00.000 --> 00:00:02.880 Hey, I'm gonna walk you through what is ChatGPT. 2 00:00:02.880 --> 00:00:06.690 This is the original AI model that everyone knows and loves. 3 00:00:06.690 --> 00:00:07.770 So you've probably heard of it, 4 00:00:07.770 --> 00:00:09.480 but we're gonna walk you through it anyway. 5 00:00:09.480 --> 00:00:11.910 What would be the next word in this sentence? 6 00:00:11.910 --> 00:00:15.090 The students open their what? 7 00:00:15.090 --> 00:00:16.110 It could be books. 8 00:00:16.110 --> 00:00:17.370 They could open their laptops, 9 00:00:17.370 --> 00:00:19.350 they could open their exams, their minds. 10 00:00:19.350 --> 00:00:20.610 There's a lot of different words 11 00:00:20.610 --> 00:00:23.430 that could come next in that sentence. 12 00:00:23.430 --> 00:00:26.250 And we can predict what word might come next 13 00:00:26.250 --> 00:00:29.670 based on the context in which we're seeing the sentence, 14 00:00:29.670 --> 00:00:31.560 and so can ChatGPT, 15 00:00:31.560 --> 00:00:34.290 and that is essentially how these models work. 16 00:00:34.290 --> 00:00:36.630 They have figured out the probability 17 00:00:36.630 --> 00:00:40.440 of what word comes next based on all of the different words 18 00:00:40.440 --> 00:00:42.360 that they've seen on the internet. 19 00:00:42.360 --> 00:00:44.850 And you know, without getting too technical, 20 00:00:44.850 --> 00:00:47.550 they have those probabilities baked into the model 21 00:00:47.550 --> 00:00:50.040 and they can decide based on the context, 22 00:00:50.040 --> 00:00:51.960 based on the prompt that you give them, 23 00:00:51.960 --> 00:00:53.250 what word comes next. 24 00:00:53.250 --> 00:00:56.070 And they specifically use tokens rather than words. 25 00:00:56.070 --> 00:00:59.130 A token is about three fourths of a word on average. 26 00:00:59.130 --> 00:01:01.350 Some differences in how they split those words up. 27 00:01:01.350 --> 00:01:03.090 But the model is trained 28 00:01:03.090 --> 00:01:05.310 until it can predict the next token. 29 00:01:05.310 --> 00:01:08.790 And then when it knows on average what the right tokens are 30 00:01:08.790 --> 00:01:10.980 to come next in the sequence, 31 00:01:10.980 --> 00:01:12.390 then you can give it a new sequence 32 00:01:12.390 --> 00:01:14.760 and figure out what comes next after that. 33 00:01:14.760 --> 00:01:16.650 And through understanding that, 34 00:01:16.650 --> 00:01:18.540 through being able to predict the next token, 35 00:01:18.540 --> 00:01:21.480 you have to understand quite a lot about how the world works 36 00:01:21.480 --> 00:01:23.760 because you need to know how things are associated 37 00:01:23.760 --> 00:01:25.620 and therefore we get the magic 38 00:01:25.620 --> 00:01:27.150 that is a large language model. 39 00:01:27.150 --> 00:01:29.970 ChatGPT specifically is based on GPT-4, 40 00:01:29.970 --> 00:01:32.250 which is a transformer model. 41 00:01:32.250 --> 00:01:33.450 Transformer models 42 00:01:33.450 --> 00:01:35.970 coming from the "Attention Is All You Need" paper 43 00:01:35.970 --> 00:01:37.080 which is a famous paper, 44 00:01:37.080 --> 00:01:39.240 but I'm not gonna run you through too much 45 00:01:39.240 --> 00:01:40.560 of the jargon here. 46 00:01:40.560 --> 00:01:41.817 It won't mean too much to you. 47 00:01:41.817 --> 00:01:44.437 You just need to know is that this is, 48 00:01:44.437 --> 00:01:46.950 you know, the current kind of state of the art 49 00:01:46.950 --> 00:01:48.480 of this type of model. 50 00:01:48.480 --> 00:01:50.580 We don't have another type of model 51 00:01:50.580 --> 00:01:53.460 other than transform models that is doing as good. 52 00:01:53.460 --> 00:01:55.110 But you know, scientists are working 53 00:01:55.110 --> 00:01:57.060 on other types of architecture as well. 54 00:01:57.060 --> 00:01:58.740 So how did they train ChatGPT? 55 00:01:58.740 --> 00:02:00.510 Well, it wasn't just the transformer model. 56 00:02:00.510 --> 00:02:02.340 That was pretty good on its own 57 00:02:02.340 --> 00:02:04.110 and that's where we had GPT-3. 58 00:02:04.110 --> 00:02:07.650 But to get ChatGPT to actually kind of make it behave 59 00:02:07.650 --> 00:02:09.900 like an assistant, we had to have additional steps. 60 00:02:09.900 --> 00:02:12.750 So after training the initial, you know, 61 00:02:12.750 --> 00:02:15.960 predict the next token model, the foundation model, 62 00:02:15.960 --> 00:02:19.110 then they had to do some fine tuning on top. 63 00:02:19.110 --> 00:02:21.570 They would have humans figure out 64 00:02:21.570 --> 00:02:25.650 which outputs were best and worst from different questions, 65 00:02:25.650 --> 00:02:27.810 and then they would use that to reward the model. 66 00:02:27.810 --> 00:02:30.720 So they would predict tokens that, you know, 67 00:02:30.720 --> 00:02:33.360 humans thought were good answers to questions. 68 00:02:33.360 --> 00:02:37.530 Then they had to develop a type of policy reward model, 69 00:02:37.530 --> 00:02:39.300 PPO reinforcement learning. 70 00:02:39.300 --> 00:02:41.190 And what that would allow you to do 71 00:02:41.190 --> 00:02:43.260 is take a prompt or a sample from the dataset, 72 00:02:43.260 --> 00:02:46.950 use that PPO model to generate the output, 73 00:02:46.950 --> 00:02:50.370 and then the reward model calculates a reward for the output 74 00:02:50.370 --> 00:02:53.490 and the reward is used to update the policy using PPO. 75 00:02:53.490 --> 00:02:55.500 So this is kind of an additional set 76 00:02:55.500 --> 00:02:57.600 of reinforcement lending on top. 77 00:02:57.600 --> 00:03:00.480 Again, technical jargon, but all that means really 78 00:03:00.480 --> 00:03:03.120 is that you have the transformer model 79 00:03:03.120 --> 00:03:04.650 that's trained on the internet, 80 00:03:04.650 --> 00:03:08.520 then you fine tune it based on what people find helpful 81 00:03:08.520 --> 00:03:10.980 and useful in terms of responses, 82 00:03:10.980 --> 00:03:13.320 and then you optimize it based on policies 83 00:03:13.320 --> 00:03:15.450 so that when you get the results, 84 00:03:15.450 --> 00:03:18.840 you can be sure that they're much better on average 85 00:03:18.840 --> 00:03:20.010 than what you would get 86 00:03:20.010 --> 00:03:22.020 if you just used the foundation model. 87 00:03:22.020 --> 00:03:23.730 It was created by OpenAI, 88 00:03:23.730 --> 00:03:27.810 You can access it at chat.openai.com or chatgpt.com. 89 00:03:27.810 --> 00:03:30.270 They have a free version that you can try out 90 00:03:30.270 --> 00:03:32.940 and the free version is pretty comprehensive and generous. 91 00:03:32.940 --> 00:03:35.040 But you can also buy plus or pro version. 92 00:03:35.040 --> 00:03:37.380 The pro version's $200 a month currently, 93 00:03:37.380 --> 00:03:39.840 and you get some additional powerful models for that. 94 00:03:39.840 --> 00:03:42.150 The features are numerous, 95 00:03:42.150 --> 00:03:43.950 and we're not gonna cover all of them here, 96 00:03:43.950 --> 00:03:47.370 but the main things that we find are useful day-to-day 97 00:03:47.370 --> 00:03:48.480 are the chat history. 98 00:03:48.480 --> 00:03:50.640 So being able to see what are the, you know, 99 00:03:50.640 --> 00:03:52.590 previous chats that you've had, 100 00:03:52.590 --> 00:03:54.030 just the ability to generate text 101 00:03:54.030 --> 00:03:56.850 so you can ask it for a blog post 102 00:03:56.850 --> 00:03:58.890 or explain how the solar system was made 103 00:03:58.890 --> 00:04:00.960 or whatever it is you're trying to generate. 104 00:04:00.960 --> 00:04:05.280 And then also image generation is built into ChatGPT 105 00:04:05.280 --> 00:04:06.113 through Dall-E 3. 106 00:04:06.113 --> 00:04:07.620 It has the ability to browse the web 107 00:04:07.620 --> 00:04:09.420 as you can get up to date current information, 108 00:04:09.420 --> 00:04:10.920 which is really powerful. 109 00:04:10.920 --> 00:04:13.260 The models themselves are a little bit confusing. 110 00:04:13.260 --> 00:04:14.370 There's lots of different models 111 00:04:14.370 --> 00:04:15.930 and they change all the time. 112 00:04:15.930 --> 00:04:18.690 But broadly speaking, we have GPT-4.0 113 00:04:18.690 --> 00:04:19.890 which is the workhorse. 114 00:04:19.890 --> 00:04:23.160 You know, the really good, useful and scalable model 115 00:04:23.160 --> 00:04:25.290 that most people use for most questions. 116 00:04:25.290 --> 00:04:27.600 We have some of the research preview-type models 117 00:04:27.600 --> 00:04:30.540 like GPT 4.5, which is a little bit more creative. 118 00:04:30.540 --> 00:04:32.250 And then you have some of the reasoning models 119 00:04:32.250 --> 00:04:35.160 like 0.1, 0.3 Mini, 0.1 Pro mode, 120 00:04:35.160 --> 00:04:36.930 and these are the ones where they think first 121 00:04:36.930 --> 00:04:38.520 before giving you a response, 122 00:04:38.520 --> 00:04:41.130 so they're pretty good for mathematical questions 123 00:04:41.130 --> 00:04:43.050 or things where you need a lot of logic. 124 00:04:43.050 --> 00:04:44.460 There's also custom GPTs, 125 00:04:44.460 --> 00:04:46.350 and you can think of these as custom prompts 126 00:04:46.350 --> 00:04:47.850 where they've given 127 00:04:47.850 --> 00:04:51.510 these versions of ChatGPT set of tools they can use. 128 00:04:51.510 --> 00:04:53.520 Maybe with the Expedia one, you could book a drip, 129 00:04:53.520 --> 00:04:56.520 but also they have a certain set of instructions 130 00:04:56.520 --> 00:04:57.900 that tells 'em how to behave. 131 00:04:57.900 --> 00:04:59.970 Things I use ChatGPT for most, 132 00:04:59.970 --> 00:05:01.410 I would say creating blog posts. 133 00:05:01.410 --> 00:05:04.770 They have a Canvas editor now, which is quite useful. 134 00:05:04.770 --> 00:05:06.000 Creating social media posts, 135 00:05:06.000 --> 00:05:08.370 particularly if you wanna create them in a specific style. 136 00:05:08.370 --> 00:05:11.040 It's very good at adhering to specific styles, 137 00:05:11.040 --> 00:05:12.900 and then automating tasks as well. 138 00:05:12.900 --> 00:05:14.760 Whenever I need to do something formal, 139 00:05:14.760 --> 00:05:17.400 fill in documentation, or do my taxes 140 00:05:17.400 --> 00:05:19.050 or appeal a parking ticket, 141 00:05:19.050 --> 00:05:22.053 then I tend to use ChatGPT for that.