1 00:00:00,080 --> 00:00:05,040 Okay, well, here we are, back in cursor, one of my very favorite places to be. 2 00:00:05,280 --> 00:00:07,160 And we're going into the week. 3 00:00:07,160 --> 00:00:08,600 One directory on the left. 4 00:00:08,640 --> 00:00:11,040 Check the LM Engineering's in block caps here. 5 00:00:11,040 --> 00:00:13,320 So you're in the right project. 6 00:00:13,320 --> 00:00:17,600 And then go to day two, which is where we find ourselves right now. 7 00:00:17,800 --> 00:00:23,120 And I need to remind you before you do anything, you have to go to the top right to where it says Select 8 00:00:23,120 --> 00:00:24,240 kernel for you. 9 00:00:24,280 --> 00:00:25,280 You click on that. 10 00:00:25,280 --> 00:00:27,240 You click in the Python Environments option. 11 00:00:27,240 --> 00:00:30,240 The first option you pick the dot VM. 12 00:00:30,440 --> 00:00:35,120 That's Python 3.12 that is matching this VM right here. 13 00:00:35,440 --> 00:00:37,920 And that means your kernel is set. 14 00:00:37,960 --> 00:00:39,800 Welcome to the day two lab. 15 00:00:40,160 --> 00:00:45,880 So I just wanted to mention here before we start that there is a page of resources on my home page and 16 00:00:45,880 --> 00:00:46,920 I've linked to it here. 17 00:00:46,960 --> 00:00:48,840 I'll try and keep adding useful resources. 18 00:00:48,840 --> 00:00:50,200 So do keep that bookmarked. 19 00:00:50,240 --> 00:00:52,360 You can go there and check it out right now if you wish. 20 00:00:52,920 --> 00:00:59,880 Okay okay so we're now going to go back again to doing what we did yesterday, but do it with a little 21 00:00:59,920 --> 00:01:00,840 bit more discipline. 22 00:01:00,840 --> 00:01:05,830 And I want to tell you about this thing called the Chat Completions API. 23 00:01:06,110 --> 00:01:12,230 And this is the name of it's the simplest way that you can call an LLM, particularly a frontier LLM. 24 00:01:12,510 --> 00:01:17,990 And it's called chat completions because it's a nod to what you're actually doing. 25 00:01:17,990 --> 00:01:25,390 When you call an LLM, you're giving it a chat a conversation so far, and you're asking it to predict 26 00:01:25,390 --> 00:01:28,350 the most likely message to come next. 27 00:01:28,470 --> 00:01:31,670 And you can think of that as as like completing a chat. 28 00:01:31,710 --> 00:01:33,510 That's what the LLM thinks it's doing. 29 00:01:33,510 --> 00:01:35,830 It's it's it's like predictive text. 30 00:01:35,830 --> 00:01:39,230 It's just trying to predict the most likely words to come next. 31 00:01:39,390 --> 00:01:44,830 And as a side effect of that, it happens to be really good at answering whatever questions it's asked. 32 00:01:44,830 --> 00:01:50,670 But all it's actually trying to do is to predict the most likely next few words, or as we'll shortly 33 00:01:50,670 --> 00:01:53,190 explain, the most likely next tokens. 34 00:01:53,590 --> 00:01:57,790 Okay, so it was actually this chat completions API. 35 00:01:57,830 --> 00:02:01,750 This this approach was invented by OpenAI first and foremost. 36 00:02:01,750 --> 00:02:08,100 But it was so popular the structure and style of this type of request became so popular that it's become 37 00:02:08,100 --> 00:02:09,060 ubiquitous. 38 00:02:09,180 --> 00:02:12,660 All of the providers offer the Chat Completions API. 39 00:02:12,820 --> 00:02:17,220 It is the kind of standard way to interact with an LLM. 40 00:02:17,300 --> 00:02:19,620 And we're going to start again with OpenAI. 41 00:02:19,660 --> 00:02:23,380 I know some of you are antsy to get off OpenAI and use free models. 42 00:02:23,580 --> 00:02:25,500 Your time is coming any second now. 43 00:02:25,540 --> 00:02:29,140 Just watch this if you don't want to use OpenAI, and we'll get to you in a second. 44 00:02:29,460 --> 00:02:30,180 All right. 45 00:02:30,300 --> 00:02:37,220 So first of all, I'm going to do a repeat of last time where we use this thing called load dot EMV, 46 00:02:37,380 --> 00:02:44,860 which is going to load in the secrets in our EMV file and check that we have OpenAI key set. 47 00:02:45,020 --> 00:02:50,780 So this load EMV function is something which which makes which loads in anything you've saved in your 48 00:02:50,780 --> 00:02:51,220 EMV. 49 00:02:51,340 --> 00:02:53,580 And then we'll just check that the key is good. 50 00:02:53,620 --> 00:02:56,500 It says API key found and it looks good so far. 51 00:02:56,500 --> 00:02:58,660 And if yours doesn't say that, then you know what to do. 52 00:02:58,700 --> 00:03:04,690 Check the EMV, check everything, and look in the look in the troubleshooting and ask me if you get 53 00:03:04,690 --> 00:03:05,170 stuck. 54 00:03:05,850 --> 00:03:08,770 Okay, let's talk about end points. 55 00:03:08,810 --> 00:03:12,610 Now, I imagine most of you know exactly what an end point is. 56 00:03:12,610 --> 00:03:14,770 It's one of those words people throw around all the time. 57 00:03:14,770 --> 00:03:15,610 But you might not. 58 00:03:15,610 --> 00:03:18,450 And for people that don't know about it, you should quickly find out. 59 00:03:18,450 --> 00:03:21,690 I explain it all in the Technical Foundations guide. 60 00:03:21,730 --> 00:03:23,890 There's nothing special at endpoint. 61 00:03:23,890 --> 00:03:31,290 It's an http url which you can call to, to to to make an API request by hitting some web address. 62 00:03:31,290 --> 00:03:34,090 And that web address would be known as an endpoint. 63 00:03:34,090 --> 00:03:35,850 It's a way that you have an API. 64 00:03:36,010 --> 00:03:40,090 But if things like API's and endpoints are new to you, then read that guide. 65 00:03:40,530 --> 00:03:41,130 Okay. 66 00:03:41,530 --> 00:03:46,370 There's an endpoint in particular which might interest you, which I want to show you right now. 67 00:03:46,370 --> 00:03:50,290 And it's an endpoint, an API endpoint offered by open AI. 68 00:03:50,490 --> 00:03:54,890 So in order to call this endpoint, which I will just do, we first have to set a couple of things which 69 00:03:54,930 --> 00:03:57,290 are very common for any HTTP request. 70 00:03:57,490 --> 00:04:01,250 There's HTTP headers, the headers that go down in that request. 71 00:04:01,290 --> 00:04:05,400 You normally you probably if you know anything about web requests. 72 00:04:05,400 --> 00:04:09,600 You specify what content type, what type of thing you want to come back. 73 00:04:09,600 --> 00:04:12,680 And that's a way of saying, I want Jason to come back, please. 74 00:04:13,000 --> 00:04:20,720 And then this is a fairly standard way of, of, of sending down a secret in an HTTP message that authorizes 75 00:04:20,720 --> 00:04:27,080 you, you put in a header called authorization and the value of that header, there's the word bearer 76 00:04:27,080 --> 00:04:31,960 and then a space and then some secret that identifies you to the third party. 77 00:04:32,120 --> 00:04:36,440 And in this case we are going to stuff the OpenAI API key. 78 00:04:36,480 --> 00:04:40,240 The thing that I got just here, we're going to stuff that into the header. 79 00:04:40,720 --> 00:04:42,880 And then I've got something here called payload. 80 00:04:42,880 --> 00:04:44,960 Payload is just a chunk of JSON. 81 00:04:45,120 --> 00:04:48,960 And it's going to be a dictionary a dictionary with two keys. 82 00:04:48,960 --> 00:04:55,320 One of them is model, which is going to be GPT five nano, the tiny version of GPT five. 83 00:04:55,480 --> 00:04:58,840 The latest model, the one of the strongest on the planet right now. 84 00:04:59,240 --> 00:05:03,480 And then the second field in this dictionary has a key of messages. 85 00:05:03,880 --> 00:05:09,350 The value is, you know, this is a list of dictionaries. 86 00:05:09,350 --> 00:05:10,630 It's a list you can see there. 87 00:05:10,630 --> 00:05:11,550 It is a list. 88 00:05:11,790 --> 00:05:19,750 Each dictionary, let's put that on another line, has a key role with value user a key content. 89 00:05:19,750 --> 00:05:24,350 And the value is tell me a fun fact okay. 90 00:05:24,390 --> 00:05:25,350 Nothing special here. 91 00:05:25,350 --> 00:05:26,390 Let's look at this. 92 00:05:26,670 --> 00:05:29,710 Uh, there is the payload JSON. 93 00:05:29,710 --> 00:05:31,150 It's is just what I said. 94 00:05:31,230 --> 00:05:35,990 Model GPT five nano messages, a list of dictionaries, role user content. 95 00:05:36,030 --> 00:05:37,550 Tell me a fun fact. 96 00:05:37,710 --> 00:05:39,470 Okay, that's a chunk of JSON. 97 00:05:39,950 --> 00:05:41,270 Uh, headers and payload. 98 00:05:41,270 --> 00:05:44,430 What we're now going to do is we're going to send that to an endpoint. 99 00:05:44,750 --> 00:05:47,110 This is the endpoint API. 100 00:05:49,590 --> 00:05:50,470 V1. 101 00:05:54,110 --> 00:05:56,230 It's chat completions. 102 00:05:56,430 --> 00:06:02,190 And sometimes when you make a Post request people think of that as like creating a new resource for 103 00:06:02,190 --> 00:06:04,230 people that are like rest, rest people. 104 00:06:04,350 --> 00:06:06,550 Uh, that's that's sometimes called creating a resource. 105 00:06:06,550 --> 00:06:12,780 So you could think of this as being like chat completions create some familiar. 106 00:06:12,900 --> 00:06:15,220 Uh, so chat completions create. 107 00:06:15,220 --> 00:06:17,780 We're passing in our headers with our API key. 108 00:06:17,780 --> 00:06:21,420 And the payload is this JSON blob right here. 109 00:06:21,580 --> 00:06:23,340 So let's run that. 110 00:06:23,700 --> 00:06:26,020 And what comes back is a bunch of JSON. 111 00:06:26,020 --> 00:06:27,780 We asked for JSON back. 112 00:06:27,820 --> 00:06:28,900 Let's see what we get back. 113 00:06:28,900 --> 00:06:30,100 Let's see what kind of JSON we get. 114 00:06:30,100 --> 00:06:30,660 Here it is. 115 00:06:30,980 --> 00:06:35,460 Uh so it's a chunk of JSON as like an ID object, various stuff. 116 00:06:35,460 --> 00:06:39,020 And then it has a field called choices and Choices. 117 00:06:39,020 --> 00:06:42,060 In this response, choices is a list. 118 00:06:42,260 --> 00:06:49,380 And the first item of that choices with index zero is something which has a field message. 119 00:06:49,980 --> 00:06:53,860 And that message is itself a JSON dictionary. 120 00:06:54,020 --> 00:06:59,060 And that has something called content, which is fun fact. 121 00:06:59,100 --> 00:07:04,340 There are possible unique games, more possible unique games of chess than there are atoms in the observable 122 00:07:04,340 --> 00:07:05,300 universe. 123 00:07:05,500 --> 00:07:09,890 About ten to the power of 120 There is a fun fact. 124 00:07:09,890 --> 00:07:15,970 And so that fun fact came back in the JSON response from our call to GPT five to that endpoint, that 125 00:07:15,970 --> 00:07:17,970 URL that I just gave you. 126 00:07:17,970 --> 00:07:20,170 And so there's obviously there's another way I could do this. 127 00:07:20,170 --> 00:07:21,530 Let me get a new code thing here. 128 00:07:21,530 --> 00:07:26,330 Let's say, uh, um, we said response JSON. 129 00:07:26,330 --> 00:07:29,450 That was the JSON that we just were just looking at right up here. 130 00:07:29,770 --> 00:07:31,170 Uh, I spelt response wrong. 131 00:07:31,330 --> 00:07:34,530 That's that's what you get if you make me type Response.json. 132 00:07:34,530 --> 00:07:38,690 So we could say, okay, let's look in the choices field. 133 00:07:39,010 --> 00:07:40,690 I guess it knows what I'm doing here. 134 00:07:40,730 --> 00:07:43,010 Look at the first element in there. 135 00:07:43,010 --> 00:07:51,290 Element index zero at the field message at the field content okay. 136 00:07:51,330 --> 00:07:51,970 You see that. 137 00:07:51,970 --> 00:07:53,370 That's you know what that's going to do. 138 00:07:53,410 --> 00:07:56,610 We're just going to choices the first one message content. 139 00:07:56,730 --> 00:07:58,250 And let's see what that prints. 140 00:07:58,370 --> 00:08:03,370 It prints that that very fun fact of course that we just looked at uh, the more possible unique games 141 00:08:03,370 --> 00:08:05,370 of chess than there are atoms in the observable universe. 142 00:08:05,930 --> 00:08:07,730 Okay, but you know what? 143 00:08:07,850 --> 00:08:14,840 Uh, so so this is a perfectly good way to call OpenAI in the cloud using an HTTP endpoint. 144 00:08:14,840 --> 00:08:15,840 And it's fine. 145 00:08:15,840 --> 00:08:16,560 We could do this. 146 00:08:16,600 --> 00:08:18,120 We could we could always type this. 147 00:08:18,480 --> 00:08:25,000 But it is kind of messy fussing around with JSON, looking in dictionaries at keys and things like that. 148 00:08:25,000 --> 00:08:31,480 It's kind of hokey, and it would be a real pain if every time we wanted to call GPT or any frontier 149 00:08:31,520 --> 00:08:39,280 model, we had to stitch together these HTTP requests, call that slash chat completions with a post 150 00:08:39,320 --> 00:08:45,160 or create and then be be sort of navigating our way through JSON objects like this. 151 00:08:45,200 --> 00:08:46,360 It would be a pain. 152 00:08:46,560 --> 00:08:48,600 It would be nice if there were a better way to do it. 153 00:08:48,640 --> 00:08:50,440 Yes, yes, I know you get the joke. 154 00:08:50,480 --> 00:08:51,600 You know what I'm gonna say? 155 00:08:51,640 --> 00:08:53,240 There is a better way to do it. 156 00:08:53,440 --> 00:08:55,800 And OpenAI created this better way. 157 00:08:55,800 --> 00:09:02,840 They made a package called OpenAI, and that package is what's known as a Python client library. 158 00:09:02,840 --> 00:09:06,760 Python client library is nothing fancy at all. 159 00:09:06,800 --> 00:09:11,550 A Python client library that you often use for APIs all over the place for you. 160 00:09:11,590 --> 00:09:15,270 Name it for APIs, for sending emails, for APIs to do so much. 161 00:09:15,470 --> 00:09:22,270 Um, and these things are typically very lightweight libraries that manufacture an HTTP request to an 162 00:09:22,270 --> 00:09:22,990 endpoint. 163 00:09:23,110 --> 00:09:28,830 And with what comes back, it turns it into Python objects so that you're not messing around with this 164 00:09:28,830 --> 00:09:33,630 kind of stuff with with these, uh, ploughing your way through JSON dictionaries. 165 00:09:33,630 --> 00:09:40,430 But you can just write some nice, elegant Python code, and that is all the OpenAI library is. 166 00:09:40,550 --> 00:09:45,790 It's a Python client library that wraps a call to an HTTP endpoint. 167 00:09:45,790 --> 00:09:47,150 It's very simple. 168 00:09:47,270 --> 00:09:48,790 It's completely open source. 169 00:09:48,790 --> 00:09:50,510 You can open it, look at it, look at all the code. 170 00:09:50,550 --> 00:09:52,150 It's perfectly simple. 171 00:09:52,470 --> 00:09:56,510 Some people, the first time they come across this, think that when you're dealing with OpenAI, the 172 00:09:56,510 --> 00:10:02,590 object and the code, that somehow we're sort of running GPT and we have some some fancy code from OpenAI. 173 00:10:02,750 --> 00:10:03,710 Not at all. 174 00:10:03,750 --> 00:10:07,910 It's vanilla code that just wraps making a web request. 175 00:10:07,910 --> 00:10:11,790 And we'll we'll quickly use that code now, but it's going to be much more familiar to you.