WEBVTT 00:00.360 --> 00:00.920 All right. 00:00.960 --> 00:02.440 Now it's the moment of truth. 00:02.600 --> 00:10.080 We now read the LinkedIn profile in as before, and hopefully this is your LinkedIn profile, not mine. 00:10.320 --> 00:13.240 And then we bring in the summary text. 00:13.280 --> 00:15.480 You should change this name to be your name. 00:15.560 --> 00:20.880 We run that and we now have our system prompt and user prompt for the LM. 00:21.240 --> 00:27.160 And the system prompt is going to be the same as last time you're acting as me or you. 00:27.200 --> 00:32.680 You're answering questions on their website, particularly related to their career, etc., etc. but 00:32.680 --> 00:34.120 there's a twist at the end here. 00:34.120 --> 00:39.840 It says if you don't know the answer to any question, use your tool to record the question you couldn't 00:39.840 --> 00:40.400 answer. 00:40.840 --> 00:46.560 And if the user is engaging in discussion, try to steer them towards getting in touch via email. 00:46.600 --> 00:48.760 Record it using your tool. 00:49.160 --> 00:55.160 Now, in theory, this isn't needed because the JSON that we've written to describe the tool already 00:55.160 --> 00:58.840 gives this kind of context and describes the tool and when it should be used. 00:59.240 --> 01:06.440 But it never hurts to be repetitive in prompting, as I'm sure you know, repetition always works well. 01:06.480 --> 01:12.800 Explain things several times over, because you will increase the probability that your model performs 01:12.800 --> 01:13.960 the way you want it to. 01:14.080 --> 01:19.840 You're biasing the model to be outputting tokens consistent with your objective, and it is important 01:19.840 --> 01:21.400 to always keep that in the back of your mind. 01:21.440 --> 01:22.750 We often like to talk about. 01:22.790 --> 01:27.190 You know, the model calls this tool and we make it sound like it almost that it's almost got something 01:27.190 --> 01:28.550 that has true autonomy. 01:28.550 --> 01:30.710 But all that's actually going on here. 01:30.750 --> 01:37.110 And always remind yourself of this from time to time, an LM is just something that's generating the 01:37.110 --> 01:41.550 most likely next tokens to follow a sequence of tokens. 01:41.550 --> 01:45.430 It's got an input and it's generating like predictive text. 01:45.470 --> 01:48.230 What is the most likely text to follow this? 01:48.670 --> 01:54.710 And if we've inserted into this prompt stuff about different tools that it could call and stuff that 01:54.710 --> 02:01.230 directs it to use them, then that just means that the most likely next tokens will be consistent with 02:01.230 --> 02:03.150 that information that we passed in. 02:03.350 --> 02:04.910 And that's all that's going on. 02:04.910 --> 02:07.470 It's about statistical most likely tokens. 02:07.470 --> 02:12.350 And that really I mean, it hurts my mind to try and think that one through, because it seems so unlikely 02:12.350 --> 02:17.750 that predicting next tokens could be so powerful as to do things like calling tools. 02:17.750 --> 02:19.950 But that really is what's happening here. 02:20.150 --> 02:22.550 All right, that was a sidebar, but it's an important sidebar. 02:22.590 --> 02:24.150 You should always try and keep that in mind. 02:24.190 --> 02:26.990 And I have a bunch of YouTube videos about that stuff. 02:26.990 --> 02:32.070 If you want to learn more about how Next token prediction is able to achieve these things. 02:32.590 --> 02:38.470 Okay, so with all of this we get to our chat function again. 02:38.590 --> 02:42.110 And this is where it really comes down to action. 02:42.110 --> 02:44.070 And we'll cover it in the next video.