WEBVTT 1 00:00:00.480 --> 00:00:01.650 Here's another example 2 00:00:01.650 --> 00:00:04.380 where these models are quite good. 3 00:00:04.380 --> 00:00:06.270 And for that, I'll reload the model, 4 00:00:06.270 --> 00:00:08.850 still the 12-billion parameters model 5 00:00:08.850 --> 00:00:12.090 with a bigger context window. 6 00:00:12.090 --> 00:00:15.390 So I'll ramp it up to, let's say, 10,000 tokens, 7 00:00:15.390 --> 00:00:18.960 just to be sure that I don't run out of context length. 8 00:00:18.960 --> 00:00:21.630 Because you can also use these models 9 00:00:21.630 --> 00:00:23.790 as content generation machines, 10 00:00:23.790 --> 00:00:27.690 especially when combining them with few-shot prompting, 11 00:00:27.690 --> 00:00:29.430 which is a prompt engineering technique, 12 00:00:29.430 --> 00:00:31.650 a quite important one. 13 00:00:31.650 --> 00:00:33.990 And you can learn all about prompt engineering 14 00:00:33.990 --> 00:00:37.530 in my complete Generative AI course, by the way. 15 00:00:37.530 --> 00:00:39.570 The idea behind few-shot prompting 16 00:00:39.570 --> 00:00:42.330 is that you provide some examples to the model 17 00:00:42.330 --> 00:00:43.890 for a given task. 18 00:00:43.890 --> 00:00:47.670 For example, I want my model to generate a LinkedIn post. 19 00:00:47.670 --> 00:00:52.670 So I will say, "You are an expert LinkedIn post generator. 20 00:00:54.360 --> 00:00:59.360 Here are some example LinkedIn posts I wrote in the past." 21 00:01:01.920 --> 00:01:05.970 And then I'll add some LinkedIn articles. 22 00:01:05.970 --> 00:01:10.970 And I use delimiters here, XML delimiters, 23 00:01:11.010 --> 00:01:13.470 which is not required technically, 24 00:01:13.470 --> 00:01:15.240 but which can help the model. 25 00:01:15.240 --> 00:01:16.740 Not just open models, 26 00:01:16.740 --> 00:01:19.020 but Large Language Models in general, 27 00:01:19.020 --> 00:01:24.020 distinguish between examples and other parts of your prompt. 28 00:01:24.750 --> 00:01:27.900 So, when having longer, more complex prompts, 29 00:01:27.900 --> 00:01:31.170 structuring your prompt by adding blocks like this 30 00:01:31.170 --> 00:01:32.640 can be useful. 31 00:01:32.640 --> 00:01:33.930 Because my idea here 32 00:01:33.930 --> 00:01:38.930 is that I wanna provide two LinkedIn article examples 33 00:01:38.940 --> 00:01:41.370 to this model here, 34 00:01:41.370 --> 00:01:43.140 to this prompt here. 35 00:01:43.140 --> 00:01:45.960 For that, I'm on my LinkedIn page here, 36 00:01:45.960 --> 00:01:48.510 and I'll simply go to my posts, 37 00:01:48.510 --> 00:01:52.173 and I'll just grab some of my more recent posts here. 38 00:01:53.190 --> 00:01:55.540 So that's the first one, I'll paste it in here. 39 00:01:57.690 --> 00:02:01.593 And then let's say I also wanna grab this one. 40 00:02:02.430 --> 00:02:04.950 So these are just some LinkedIn articles 41 00:02:04.950 --> 00:02:06.450 I wrote in the past, 42 00:02:06.450 --> 00:02:10.320 and I'll add them between those XML tags here 43 00:02:10.320 --> 00:02:14.699 to clearly highlight the places where one post starts 44 00:02:14.699 --> 00:02:17.430 and then ends and the next one starts. 45 00:02:17.430 --> 00:02:20.257 And then after this second example, I'll say, 46 00:02:20.257 --> 00:02:25.013 "Above, you see two example posts, or LinkedIn posts, 47 00:02:26.310 --> 00:02:28.473 I wrote in the past. 48 00:02:29.550 --> 00:02:34.550 Use the same writing style as shown in those posts, 49 00:02:35.190 --> 00:02:38.580 but don't use the content. 50 00:02:38.580 --> 00:02:43.580 Instead, using the above shown writing style, 51 00:02:43.920 --> 00:02:45.540 generate a new... 52 00:02:45.540 --> 00:02:49.270 Generate a new LinkedIn post 53 00:02:50.400 --> 00:02:55.143 about using open LLMs locally. 54 00:02:56.580 --> 00:02:57.750 The post should cover 55 00:02:57.750 --> 00:03:00.810 the following core concepts and topics." 56 00:03:00.810 --> 00:03:03.090 And then I created a list of concepts 57 00:03:03.090 --> 00:03:04.683 I want the post to cover. 58 00:03:06.300 --> 00:03:08.550 Now, I could, of course, refine that, 59 00:03:08.550 --> 00:03:10.350 add more instructions to this prompt, 60 00:03:10.350 --> 00:03:11.910 but I can also now send this 61 00:03:11.910 --> 00:03:15.780 to this 12-billion parameter Gemma 3 model. 62 00:03:15.780 --> 00:03:19.830 So not even the most capable or powerful Gemma 3 model. 63 00:03:19.830 --> 00:03:23.400 I could also run the 27-billion parameters model here, 64 00:03:23.400 --> 00:03:24.840 but I'm using this one, 65 00:03:24.840 --> 00:03:28.860 and I'm sending this as a few-shot prompt to this model 66 00:03:28.860 --> 00:03:31.230 because I included some examples. 67 00:03:31.230 --> 00:03:34.350 And because of these examples I included here, 68 00:03:34.350 --> 00:03:37.080 I am getting a LinkedIn post, 69 00:03:37.080 --> 00:03:41.250 which I could now take and tweak and then share on LinkedIn. 70 00:03:41.250 --> 00:03:45.150 I'm getting that post from that locally running model. 71 00:03:45.150 --> 00:03:49.440 So without using any paid or proprietary model, 72 00:03:49.440 --> 00:03:54.060 and to me, on first sight, this post doesn't look too bad. 73 00:03:54.060 --> 00:03:56.460 If I would've included more examples here 74 00:03:56.460 --> 00:04:00.450 and/or use the more capable model, it might be even better. 75 00:04:00.450 --> 00:04:02.490 And just to show you the difference I get 76 00:04:02.490 --> 00:04:05.730 from including these examples, 77 00:04:05.730 --> 00:04:10.730 let me grab that same prompt without the examples. 78 00:04:12.360 --> 00:04:15.127 I'll open a new chat, insert it here, and say, 79 00:04:15.127 --> 00:04:17.130 "Generate a new LinkedIn post 80 00:04:17.130 --> 00:04:18.780 about using open LLMs locally." 81 00:04:18.780 --> 00:04:23.780 And I'll say, "You are an expert LinkedIn post generator." 82 00:04:26.010 --> 00:04:28.920 So now, I have no examples in there. 83 00:04:28.920 --> 00:04:31.480 I'll bring back my bullets here 84 00:04:32.460 --> 00:04:34.623 and I'll send this to that same model. 85 00:04:35.640 --> 00:04:36.660 And as you see, 86 00:04:36.660 --> 00:04:40.710 I get quite a different kind of post with way more emojis, 87 00:04:40.710 --> 00:04:43.530 as these Large Language Models tend to do. 88 00:04:43.530 --> 00:04:45.570 They tend to include more emojis. 89 00:04:45.570 --> 00:04:47.400 I also got hashtags here, 90 00:04:47.400 --> 00:04:51.093 even though I don't want to have hashtags in my posts. 91 00:04:52.170 --> 00:04:55.830 So, you see these posts are quite different 92 00:04:55.830 --> 00:04:58.080 from the posts I got before. 93 00:04:58.080 --> 00:04:59.280 Here they are again, 94 00:04:59.280 --> 00:05:01.410 these are the posts I got 95 00:05:01.410 --> 00:05:04.320 with my extra examples that I added. 96 00:05:04.320 --> 00:05:05.250 And this shows you 97 00:05:05.250 --> 00:05:07.800 that for tasks that involve few-shot prompting, 98 00:05:07.800 --> 00:05:10.650 like for example here, content generation, 99 00:05:10.650 --> 00:05:13.230 locally running models can be amazing 100 00:05:13.230 --> 00:05:14.970 and can be all you need, 101 00:05:14.970 --> 00:05:17.040 and can therefore also replace 102 00:05:17.040 --> 00:05:19.113 proprietary models potentially.