1 00:00:00,120 --> 00:00:06,160 Yes, I know that the week is already finished, but I wanted to slip in a little bit extra something. 2 00:00:06,160 --> 00:00:06,400 Just. 3 00:00:06,400 --> 00:00:07,400 Just a nugget. 4 00:00:07,400 --> 00:00:09,120 Something that I think will entertain you. 5 00:00:09,120 --> 00:00:13,000 It will only take a couple of minutes and I think you'll agree it'll be worth it. 6 00:00:13,040 --> 00:00:14,160 Please humor me. 7 00:00:14,200 --> 00:00:19,640 Come on in with me to week two and to this new notebook. 8 00:00:19,880 --> 00:00:21,480 Extra, extra. 9 00:00:21,520 --> 00:00:22,600 Did you see it hiding there? 10 00:00:22,600 --> 00:00:26,080 This is something that's a special extra that I want to go through with you. 11 00:00:26,280 --> 00:00:31,440 This is something inspired by the tech writer Simon Willison, who has a wonderful blog. 12 00:00:31,640 --> 00:00:37,640 And one of the things that he likes to do is put models to the test by asking them to draw pelicans 13 00:00:37,640 --> 00:00:39,880 riding on bikes and some other things. 14 00:00:39,880 --> 00:00:46,120 And we're going to do something similar now just to explore a bunch of different models, doing a task 15 00:00:46,280 --> 00:00:51,880 as a as a preview, as a teaser for what's to come in the future weeks, including particularly week 16 00:00:51,920 --> 00:00:54,920 four when we pick the right model for a task at hand. 17 00:00:55,240 --> 00:00:56,480 All right, let's get into it. 18 00:00:56,520 --> 00:01:03,120 We are going to ask a bunch of models to come up with a picture, draw an image, a silly image to test 19 00:01:03,120 --> 00:01:03,680 them out. 20 00:01:03,680 --> 00:01:05,740 But here's the thing to keep in mind. 21 00:01:05,780 --> 00:01:08,580 We've already seen how to use image generation. 22 00:01:08,580 --> 00:01:11,580 We've already seen how to drive a model to actually draw. 23 00:01:11,820 --> 00:01:13,220 Create an image. 24 00:01:13,220 --> 00:01:16,300 And you may know that it uses these diffusion model things. 25 00:01:16,300 --> 00:01:19,020 We're actually going to use one in week three coming right up. 26 00:01:19,460 --> 00:01:21,620 But we're going to do something different. 27 00:01:21,620 --> 00:01:28,540 We're not going to ask it to generate an image, but rather we're going to ask it to draw an SVG. 28 00:01:28,940 --> 00:01:35,060 And SVG is a type of image file, but in fact what it contains within it is just a straight up description 29 00:01:35,340 --> 00:01:42,460 in in like in letters of how to draw the lines and shapes that make up an image, it's literally like 30 00:01:42,460 --> 00:01:49,100 draw a line here, a box here, fill it in, draw a box here, laid out as a series of steps in a in 31 00:01:49,100 --> 00:01:51,700 a document that's like an XML style document. 32 00:01:51,940 --> 00:01:53,780 That's what an SVG is. 33 00:01:53,780 --> 00:02:00,740 So in many ways, asking a model to generate an SVG is actually asking it to think through. 34 00:02:00,860 --> 00:02:06,300 How would it go about constructing this image and drawing it line by line. 35 00:02:06,300 --> 00:02:12,830 And that's really interesting because it's more about a model's intelligence of being able to to craft 36 00:02:12,830 --> 00:02:17,910 this painting than it is about image generation, which is more about using special kinds of models 37 00:02:17,910 --> 00:02:20,350 that are good based on all of their training data. 38 00:02:20,590 --> 00:02:21,070 Okay. 39 00:02:21,310 --> 00:02:22,310 That's the backdrop. 40 00:02:22,350 --> 00:02:24,630 The other piece of backdrop is about open router. 41 00:02:24,670 --> 00:02:30,110 I know I've mentioned open router before, but open router is a platform that gives you access to many 42 00:02:30,110 --> 00:02:30,870 models. 43 00:02:30,870 --> 00:02:34,070 And it does so by routing your request to the different models. 44 00:02:34,070 --> 00:02:39,310 And one of the great innovations about Open Router is that you can use one API key, your key to open 45 00:02:39,310 --> 00:02:43,070 router themselves and use that as your sort of gateway in. 46 00:02:43,230 --> 00:02:46,510 It is of course, it's it's not an abstraction layer. 47 00:02:46,510 --> 00:02:47,390 It's a router. 48 00:02:47,430 --> 00:02:48,950 We talked about the difference between them. 49 00:02:48,950 --> 00:02:51,030 So I'm going to use open router. 50 00:02:51,070 --> 00:02:52,870 You can either just watch what I do. 51 00:02:52,870 --> 00:02:54,630 You don't need to run this if you don't want to. 52 00:02:54,670 --> 00:02:58,070 But if you'd also like to do this then you should do it too. 53 00:02:58,190 --> 00:03:01,070 So I've got some imports as usual. 54 00:03:01,070 --> 00:03:07,710 And here you remember from guide nine in the guides folder that we can use open router base URL. 55 00:03:07,750 --> 00:03:09,310 We can take an API key. 56 00:03:09,350 --> 00:03:12,830 I'll just check that the API key is set up and it looks good so far. 57 00:03:12,870 --> 00:03:14,390 Hopefully you got that. 58 00:03:14,490 --> 00:03:20,690 And now I'm creating an instance of the Python client library that's configured for open router. 59 00:03:21,010 --> 00:03:21,610 Okay. 60 00:03:21,770 --> 00:03:23,450 And now for the assignment. 61 00:03:23,450 --> 00:03:26,890 And I don't mean an assignment for you, I mean an assignment for the model. 62 00:03:27,010 --> 00:03:29,330 So we are going to come up with a challenge. 63 00:03:29,330 --> 00:03:33,890 And in my case I'm going with a panda rollerblading to work. 64 00:03:34,050 --> 00:03:36,850 But you should come up with something different and interesting. 65 00:03:36,890 --> 00:03:37,890 Not too complicated. 66 00:03:37,890 --> 00:03:42,490 You want to give it some creative freedom and something that it could realistically do out of lines 67 00:03:42,490 --> 00:03:43,170 and shapes. 68 00:03:43,570 --> 00:03:49,770 So the prompt generate an SVG of this challenge and respond with the SVG only. 69 00:03:49,970 --> 00:03:50,970 That's the setup. 70 00:03:50,970 --> 00:03:56,650 And then I have a function artist that takes any model we want to give it, and it calls open router, 71 00:03:58,130 --> 00:04:03,450 create passes in the model, the messages and the reasoning effort. 72 00:04:03,730 --> 00:04:09,130 And it times it and it returns the sum summary of it and the results. 73 00:04:09,170 --> 00:04:11,370 That's the function artist. 74 00:04:11,450 --> 00:04:12,050 Okay. 75 00:04:12,330 --> 00:04:14,130 Now let's look at the models. 76 00:04:14,130 --> 00:04:21,430 So we're going to call artist repeatedly for a bunch of models for the open source version of GPT 120, 77 00:04:21,470 --> 00:04:25,190 be the GPT five nano with low reasoning. 78 00:04:25,230 --> 00:04:27,070 We're getting stronger and stronger, I think. 79 00:04:27,270 --> 00:04:29,150 Deep seek the latest version for me. 80 00:04:29,190 --> 00:04:36,150 3.2 Kimi Katou thinking is as of for me right now, beginning of 2026. 81 00:04:36,470 --> 00:04:42,230 The strongest open source model on the planet, I do believe may be different for you, and you can 82 00:04:42,230 --> 00:04:43,590 feel free to pick whatever you want. 83 00:04:43,630 --> 00:04:45,070 This is from moonshot AI. 84 00:04:45,110 --> 00:04:50,630 Chinese AI startup then grok 4.1 fast from Mr. Musk, then Claude. 85 00:04:50,630 --> 00:04:57,510 Opus 4.5 the top model from anthropic right now GPT 5.2 with high reasoning. 86 00:04:57,510 --> 00:04:59,670 The strongest model from OpenAI for me right now. 87 00:04:59,950 --> 00:05:02,630 And finally, Gemini three Pro preview. 88 00:05:02,910 --> 00:05:04,350 These are our models. 89 00:05:04,350 --> 00:05:06,990 And when I run this off it goes. 90 00:05:06,990 --> 00:05:08,590 They are all going to run. 91 00:05:08,590 --> 00:05:13,190 Now they are running one after another, doing the processing and storing back the results. 92 00:05:13,190 --> 00:05:18,110 And for anyone that's taken the Agentic track or some of my other courses, you'll know well how this 93 00:05:18,110 --> 00:05:23,350 could be done as async code, which would mean that they would actually run in parallel rather than 94 00:05:23,480 --> 00:05:26,280 one by one, but I didn't want to overcomplicate by putting it here. 95 00:05:26,400 --> 00:05:30,500 So because I've done it like this, this simple, straightforward way, it's going to take like five 96 00:05:30,500 --> 00:05:31,600 minutes for this to run. 97 00:05:31,600 --> 00:05:36,400 But for you, if you if you know it, feel free to turn this into something that is async and it will 98 00:05:36,400 --> 00:05:43,040 take just just a minute or two, but I will see you in five or so minutes when all of our artists have 99 00:05:43,040 --> 00:05:43,760 wrapped up. 100 00:05:43,800 --> 00:05:45,840 Okay, that did not take five minutes. 101 00:05:45,840 --> 00:05:47,720 That took 12.5 minutes. 102 00:05:47,880 --> 00:05:49,960 I'm sorry if it took that long for you to once. 103 00:05:49,960 --> 00:05:50,800 It took me five minutes. 104 00:05:50,800 --> 00:05:55,600 I think there's quite a variability, but it's time for us now to view the results. 105 00:05:55,600 --> 00:05:57,360 It's going to draw each one. 106 00:05:57,360 --> 00:06:02,680 I wrote a little script called revealer, which just means that it paints the SVG line by line so we 107 00:06:02,680 --> 00:06:03,840 can see it coming. 108 00:06:03,880 --> 00:06:07,240 Let's now run this to see model by model. 109 00:06:07,240 --> 00:06:10,640 What kind of drawing did they come up with for our panda? 110 00:06:10,680 --> 00:06:13,200 Okay, I've zoomed out a bit and now we're off. 111 00:06:13,400 --> 00:06:14,400 This is the first one. 112 00:06:14,400 --> 00:06:16,960 This is GPT OS 120 b. 113 00:06:17,240 --> 00:06:18,360 That looks pretty good. 114 00:06:18,360 --> 00:06:19,280 I quite like that. 115 00:06:19,280 --> 00:06:20,240 Nice rollerblades. 116 00:06:20,240 --> 00:06:20,920 Well done. 117 00:06:20,960 --> 00:06:23,720 An open source model did a fine job. 118 00:06:24,120 --> 00:06:24,640 All right. 119 00:06:24,680 --> 00:06:26,840 And after a bit of a pause, this is the next one. 120 00:06:26,840 --> 00:06:33,060 This is deep seek 3.2 ha with, uh, rollerblades at different heights. 121 00:06:33,100 --> 00:06:34,140 There's the road. 122 00:06:34,420 --> 00:06:37,980 A nice and big diagram it's given us that is deep seek. 123 00:06:38,180 --> 00:06:43,300 Here is now Kimi Katou, the strongest open source model on the planet. 124 00:06:43,300 --> 00:06:45,900 And wearing a hat or a speech bubble. 125 00:06:45,900 --> 00:06:46,740 I'm not sure what that is. 126 00:06:46,740 --> 00:06:47,780 Maybe you could tell what that is. 127 00:06:47,780 --> 00:06:48,620 I'm not sure. 128 00:06:48,820 --> 00:06:49,660 Seems interesting. 129 00:06:49,660 --> 00:06:53,580 An interesting interpretation here is Grock 4.1 fast. 130 00:06:53,740 --> 00:06:54,940 And that's fun. 131 00:06:55,220 --> 00:06:56,020 That's fun. 132 00:06:56,180 --> 00:06:58,300 Uh, again, some hat situation going on. 133 00:06:58,340 --> 00:07:00,300 Work written assertively here. 134 00:07:00,420 --> 00:07:02,460 And rollerblades for sure. 135 00:07:02,860 --> 00:07:07,580 This is Claude, opus 4.5 Anthropic's strongest model as of today. 136 00:07:07,620 --> 00:07:08,700 Quite impressive. 137 00:07:08,700 --> 00:07:14,220 Rollerblades, little action lines and the office written as office in case we were in confusion. 138 00:07:14,500 --> 00:07:18,700 And now GPT 5.2, the strongest model from OpenAI. 139 00:07:19,140 --> 00:07:20,700 And look at this. 140 00:07:20,940 --> 00:07:22,500 Look at this Panda. 141 00:07:22,700 --> 00:07:25,900 And there's like action lines coming I love it. 142 00:07:25,900 --> 00:07:27,380 Swirly action lines. 143 00:07:27,420 --> 00:07:30,460 Last but not least, Gemini three Pro. 144 00:07:30,700 --> 00:07:33,990 Oh, wow Wow, look at that. 145 00:07:33,990 --> 00:07:37,190 Look at the flailing rollerblades in the action lines. 146 00:07:37,390 --> 00:07:38,790 That's quite an image. 147 00:07:39,270 --> 00:07:41,070 That is quite an image. 148 00:07:41,390 --> 00:07:46,550 Okay, I would say Gemini three Pro for the win. 149 00:07:47,310 --> 00:07:49,510 That's an impressive bunch of images. 150 00:07:49,510 --> 00:07:50,910 I absolutely love it. 151 00:07:50,910 --> 00:07:54,030 So GPT five nano was the only one that wasn't able to make it. 152 00:07:54,070 --> 00:07:57,710 It had a problem generating, but everybody else did a fine job. 153 00:07:57,870 --> 00:08:01,910 A great set of drawn svg's, line by line. 154 00:08:02,150 --> 00:08:02,470 Uh. 155 00:08:02,510 --> 00:08:03,590 Really impressive. 156 00:08:03,630 --> 00:08:05,190 Hopefully you twisted it. 157 00:08:05,190 --> 00:08:06,270 You came up with something different. 158 00:08:06,270 --> 00:08:12,630 You've got a great set of images, and you're enjoying this hands on experience, exploring the capabilities 159 00:08:12,630 --> 00:08:17,470 of frontier models and open source models through the Open Router API. 160 00:08:17,790 --> 00:08:21,670 And that now, finally, really is the conclusion of week two. 161 00:08:21,910 --> 00:08:23,710 That's the end of the extra. 162 00:08:23,710 --> 00:08:25,070 I hope you loved it. 163 00:08:25,110 --> 00:08:31,310 Hope this left you excited about the real scientific ways coming up in week four, but for now, it's 164 00:08:31,310 --> 00:08:36,550 time to take a breather and then go in to week three for open source models. 165 00:08:36,550 --> 00:08:37,870 I can't wait to see you then.