WEBVTT 00:00.320 --> 00:02.360 So here we are, back in cursor. 00:02.400 --> 00:03.680 A great place to be. 00:03.960 --> 00:06.560 So on the left is our directory structure. 00:06.560 --> 00:12.800 I open up the foundations folder and I'm going to go to lab two which is what we're doing right now. 00:12.800 --> 00:14.360 Welcome to the second lab week one. 00:14.360 --> 00:15.160 Day three. 00:15.440 --> 00:17.720 We're going to work with a lot of models. 00:18.080 --> 00:20.920 Now let me start with just some quick points. 00:21.280 --> 00:26.000 First of all, I want to mention that the way that I like to do this, the way that that I collaborate 00:26.000 --> 00:29.360 with you may be different to some other courses that you've taken. 00:29.360 --> 00:31.200 And it's something that is a bit controversial. 00:31.200 --> 00:36.600 Some people like it, some people don't like it, and I hope that you'll get used to it. 00:36.960 --> 00:42.800 Typically, I don't like to just sit here typing code because I feel like that slows down the momentum. 00:43.000 --> 00:49.840 Rather, I look at cells, I explain what they do, I run them, I inspect them, I print them, and 00:49.840 --> 00:51.760 that's what I recommend you do as well. 00:52.080 --> 00:57.280 Watch while I go through this and explain it, and then come back afterwards and run it yourself. 00:57.280 --> 00:58.920 And then add in print statements. 00:58.960 --> 00:59.940 Change things. 00:59.940 --> 01:01.020 Experiment. 01:01.020 --> 01:04.140 It's all about experimentation and that's what you should be doing. 01:04.140 --> 01:09.460 So for people that would prefer it if I, if I sat here typing, I'm sorry, I will try and do it from 01:09.460 --> 01:09.820 time to time. 01:09.820 --> 01:10.940 I'm going to do it right here. 01:10.940 --> 01:15.300 I will do it occasionally, but most of the time I'm going to explain what I'm doing and I hope you'll 01:15.300 --> 01:15.940 get used to that. 01:15.940 --> 01:19.620 And you'll see that there are pros and cons of everything, but that's the way I like to do it. 01:19.980 --> 01:25.260 Now, one thing I will really, really encourage you is written here that that once you've gone through 01:25.260 --> 01:30.980 this and made some changes, done the exercises, perhaps it would be fabulous if you were willing to 01:31.020 --> 01:38.220 save your changes in the folder called Community Contributions and raise a PR or a pull request for 01:38.220 --> 01:43.100 me to merge that into the repo so that other students get to benefit from this as well. 01:43.140 --> 01:47.580 To my great joy for my other course, there have been hundreds of PRS. 01:47.620 --> 01:53.300 There are tons of examples from people like you taking the course and submitting what they do with all 01:53.340 --> 01:57.180 sorts of fun and interesting examples, and it's been a real joy to see. 01:57.220 --> 02:02.640 And it's also been valuable for other students to experiment with, with people's contributions. 02:02.640 --> 02:04.520 So it's a fabulous thing to do. 02:04.560 --> 02:05.960 I really encourage it. 02:06.200 --> 02:11.880 In addition to that, something which is a good career thing to try is it's good if you have your own 02:11.880 --> 02:17.240 GitHub repo, and as you work on different projects, put some of your own stuff there as well. 02:17.280 --> 02:19.480 Showcase the work that you're doing. 02:19.600 --> 02:24.240 And if you put a post on LinkedIn to talk about some of the projects that you've worked on or some of 02:24.280 --> 02:28.600 the things you've done or what you've learned from it, and you tag me on it so that I get a notification 02:28.600 --> 02:29.480 in LinkedIn. 02:29.520 --> 02:34.040 Then when I see that, I'll come in and I'll weigh in and I'll make sure that I post something there 02:34.040 --> 02:34.760 as well. 02:34.800 --> 02:36.840 And that amplifies what you're doing. 02:36.840 --> 02:41.080 It means that it goes out to all the other people I'm connected with on this course, and people can 02:41.120 --> 02:42.760 then weigh in with their thoughts too. 02:42.800 --> 02:48.520 And it helps to draw attention to this, the skills you're acquiring, your expertise and your examples. 02:48.520 --> 02:52.960 And that's the kind of thing that that might attract the attention, perhaps, of a future client of 02:52.960 --> 02:58.580 yours, someone that might contact you for some business and potentially even a future employer as well. 02:58.580 --> 02:59.980 So it's a great thing to do. 02:59.980 --> 03:04.060 I strongly encourage it, and I'll be here to amplify the work that you do. 03:04.980 --> 03:06.180 Okay, let's get on with it. 03:06.180 --> 03:11.220 So it begins, of course, with some inputs and I press shift Enter to run these inputs. 03:11.260 --> 03:13.820 Be sure to run the inputs or you will get name errors. 03:14.140 --> 03:14.660 Okay. 03:14.780 --> 03:17.540 So always remember to do this as the comment what is that. 03:17.580 --> 03:22.820 Well it is of course doing a load to bring in the environment variables. 03:22.820 --> 03:28.780 And you remember that I like to have override is true to avoid any existing environment variables taking 03:28.780 --> 03:29.500 priority. 03:29.500 --> 03:30.660 So that has happened. 03:30.780 --> 03:35.940 So now that that's run, what we should find is that we've got a bunch of environment variables. 03:35.980 --> 03:40.900 Now I've been I've been quite greedy and I've set up a bunch of different APIs. 03:40.900 --> 03:45.940 And I don't expect you to do the same, but should you wish to, here are all of the APIs that you could 03:45.980 --> 03:46.620 set up. 03:46.620 --> 03:48.340 And I have done so and I have enjoyed it. 03:48.620 --> 03:53.580 Open AI, anthropic, Google's Gemini, Deep Seek, and grok. 03:53.940 --> 03:59.520 It's worth mentioning grok as when I did it at least doesn't have like an upfront a minimum. 03:59.520 --> 04:03.960 You only pay for what you use and it's really cheap, so this is a good one to try. 04:04.160 --> 04:06.760 Deep seek has a $2 upfront. 04:06.800 --> 04:08.160 At least it did for me. 04:08.200 --> 04:12.080 And then you draw down against that $2 and I haven't been able to spend $2. 04:12.720 --> 04:16.680 Gemini is free for within certain tiering. 04:17.000 --> 04:22.800 And anthropic and OpenAI both require an upfront, which I believe is $5 in the US for both. 04:22.800 --> 04:26.920 Although OpenAI appears to have a new deal right now, which I don't know how long that will last for, 04:26.920 --> 04:28.360 but you could always go and check it out. 04:28.840 --> 04:34.200 So this code here will load in the environment variables and just check that these keys look right if 04:34.240 --> 04:35.080 you're using them. 04:35.080 --> 04:38.120 And when I run this it gets this print here. 04:38.120 --> 04:43.360 And you can see that I've got keys that look like they should look or with the right kinds of prefixes. 04:43.520 --> 04:47.560 And if you run this cell and you haven't set up some of these keys, you'll get you'll get some sort 04:47.560 --> 04:48.800 of not set. 04:48.800 --> 04:50.360 But don't worry about that. 04:50.360 --> 04:53.760 Just don't don't call that particular model okay. 04:54.210 --> 04:58.530 And now to get started, we're going to make a bunch of different calls to LMS. 04:58.690 --> 05:00.890 And there's two objectives for this. 05:00.890 --> 05:06.370 One of them is just to show you the different kinds of API's, the different style of messaging that 05:06.370 --> 05:08.490 we have with with different LMS. 05:08.490 --> 05:13.090 So you get a sense of it, because we'll be talking to lots of LMS over the next few weeks, and the 05:13.090 --> 05:17.170 other is to do some orchestration between models. 05:17.290 --> 05:21.810 A few of the design patterns that we talked about last time to see how that works out. 05:22.050 --> 05:25.290 So the first thing we're going to do here is we have a question. 05:25.330 --> 05:26.890 We're going to have a request. 05:26.930 --> 05:32.970 Please come up with a challenging, nuanced question that I can ask a number of LMS to evaluate their 05:32.970 --> 05:34.010 intelligence. 05:34.330 --> 05:36.130 Please answer only with the question. 05:36.170 --> 05:37.650 No explanation. 05:37.650 --> 05:41.850 So that's going in a variable called request and into a variable called messages. 05:41.850 --> 05:46.970 We're going to have this standard OpenAI construct role is user. 05:47.130 --> 05:51.410 The content is going to be this request that we have right here. 05:51.690 --> 05:55.790 So in fact what I'm going to do is I'm going to add another code cell here. 05:55.790 --> 05:56.910 Let's run this. 05:57.590 --> 05:59.390 And I'm just going to print messages. 05:59.390 --> 06:03.150 Just so you see this, just in case it's not completely clear to you. 06:03.150 --> 06:06.510 It's a list of dictionaries role user content. 06:06.510 --> 06:07.710 And there's the content. 06:07.950 --> 06:11.590 Now, people who've been on my prior course or generally know about these things might be wondering 06:11.590 --> 06:13.230 why there isn't a system message. 06:13.270 --> 06:17.830 Well, we often do do talk about the need for a system message and a user message, but system messages 06:17.870 --> 06:18.910 are optional. 06:18.910 --> 06:21.750 You don't need one, particularly if it is a kind of standard. 06:21.790 --> 06:23.470 You are a helpful assistant then. 06:23.470 --> 06:24.430 Then you don't need to say it. 06:24.430 --> 06:27.710 It will work perfectly well without it, as you will see. 06:28.110 --> 06:29.190 Let's give it a try. 06:29.230 --> 06:35.790 We're going to to use the same API structure that we did last time that hopefully most of you are pretty 06:35.790 --> 06:39.030 familiar with, and that cursor will just fill in for you if you start typing this. 06:39.030 --> 06:48.470 Anyway, we create a new instance of the OpenAI Python client library and we call OpenAI create. 06:48.510 --> 06:52.810 We pass in the model, we're going to use the cheap GPT four mini. 06:52.850 --> 07:00.570 We pass in this list of dicts, and we get back and we ask for the choices zero message content, and 07:00.570 --> 07:03.650 we'll print it, and we're putting it in a variable called question. 07:03.650 --> 07:05.330 Let's see what we get. 07:05.930 --> 07:14.010 So how would you analyze the ethical implications of using AI in predictive policing? 07:14.250 --> 07:14.530 Wow. 07:14.570 --> 07:18.970 Considering factors such as bias, accountability and societal impact that is a good question. 07:19.010 --> 07:19.810 Goodness. 07:19.850 --> 07:20.290 Wow. 07:20.410 --> 07:23.050 Okay, that's a hard question for sure. 07:23.210 --> 07:25.290 This is going to be an interesting challenge. 07:25.290 --> 07:28.650 We will see what happens next okay. 07:28.970 --> 07:36.970 So what I'm going to do next is I'm going to first set up an empty list called Competitors List that 07:36.970 --> 07:39.810 we will be filling up with different names of competitors. 07:39.810 --> 07:42.890 And I'm then going to have another list which I'm going to call answers. 07:43.210 --> 07:49.450 And that's going to be where we will fill up different answers from different llms answering this question. 07:49.910 --> 07:54.270 and then we will put together the, the messages that we're going to to, to say to them. 07:54.510 --> 07:57.830 And I love the way Castor does this, but this is exactly what we want. 07:57.870 --> 08:00.270 We want a list of Dick's role is user. 08:00.310 --> 08:06.670 The content will be this question, the exact question that GPT four mini just asked us. 08:07.270 --> 08:11.390 Okay, we're now going to go through and I need to remember to run that cell. 08:11.750 --> 08:15.470 We'll run that cell and we will go through and ask some models. 08:15.590 --> 08:18.590 And why not start with the same model that came up with the question? 08:18.590 --> 08:22.710 We will start by asking GPT four or mini its own question. 08:22.710 --> 08:23.790 So this is what we do. 08:23.790 --> 08:27.790 We're going to have a variable model name which we will put GPT four or mini in. 08:27.950 --> 08:33.870 And then we will say response is OpenAI create that standard API. 08:34.030 --> 08:35.270 We pass in the model name. 08:35.270 --> 08:37.590 We pass in the messages that we've got right here. 08:37.910 --> 08:41.590 And then we take back the answer choice is zero message content. 08:41.710 --> 08:43.670 We will print that answer. 08:43.710 --> 08:45.270 Let's do better than than print it. 08:45.270 --> 08:47.350 Let's use markdown instead. 08:47.350 --> 08:52.010 So we'll say display markdown answer so that we get to see it in a nice format. 08:52.010 --> 08:54.370 Because, you know, these models love to respond in markdown. 08:54.690 --> 08:58.730 And then we will add GPT four or mini to our competitors. 08:58.730 --> 09:02.330 And we will add that answer to our list of answers. 09:03.050 --> 09:03.690 Sound good? 09:03.930 --> 09:04.970 Let's give this a try. 09:05.490 --> 09:07.210 So it's running while it's running. 09:07.490 --> 09:08.290 Take a look. 09:08.330 --> 09:14.610 And and whilst you should never need to memorize these APIs because it will always fill it in for you, 09:14.610 --> 09:18.330 it's good to have an instinct for what this API looks like. 09:18.330 --> 09:19.770 The chat completions API. 09:20.410 --> 09:22.610 All right, so here came the response. 09:22.650 --> 09:25.370 It was of course in beautiful markdown. 09:25.370 --> 09:28.250 And you can see it's a I am not going to read it through now. 09:28.290 --> 09:29.010 Live with you. 09:29.010 --> 09:35.050 But it looks like it's a pretty robust answer that has good categories and a bunch of different sections 09:35.050 --> 09:36.130 and a conclusion. 09:36.370 --> 09:40.050 The ethical implications are complex and multifaceted. 09:40.250 --> 09:40.890 Okay. 09:41.090 --> 09:42.010 Fair enough. 09:42.250 --> 09:42.810 That's good. 09:42.810 --> 09:44.970 That's GPT four mini. 09:45.010 --> 09:47.330 We will now move on to another API.