WEBVTT 1 00:00:00.000 --> 00:00:02.190 Hey, in this video, you're gonna have a look and explore 2 00:00:02.190 --> 00:00:04.770 what is this concept called hallucinations. 3 00:00:04.770 --> 00:00:06.600 We'll look at what hallucinations are 4 00:00:06.600 --> 00:00:07.860 and how you can avoid them 5 00:00:07.860 --> 00:00:10.740 when you're using large language models such as ChatGPT. 6 00:00:10.740 --> 00:00:13.530 Firstly, hallucinations are basically 7 00:00:13.530 --> 00:00:16.920 some type of factually incorrect text 8 00:00:16.920 --> 00:00:18.690 that has been created by the LLM 9 00:00:18.690 --> 00:00:19.770 or the large language model. 10 00:00:19.770 --> 00:00:22.890 Large language models are simply predicting the next token 11 00:00:22.890 --> 00:00:24.810 and therefore they can often produce 12 00:00:24.810 --> 00:00:26.850 factually incorrect text. 13 00:00:26.850 --> 00:00:29.130 There was a study done in 2022 14 00:00:29.130 --> 00:00:32.880 about ChatGPT was creating non-factual references 15 00:00:32.880 --> 00:00:34.410 within the science community. 16 00:00:34.410 --> 00:00:36.180 And this can be a really large problem 17 00:00:36.180 --> 00:00:38.490 in areas like law and healthcare. 18 00:00:38.490 --> 00:00:41.610 So let's have a look at an example inside of ChatGPT 19 00:00:41.610 --> 00:00:43.650 where we're telling it to create some fake 20 00:00:43.650 --> 00:00:45.570 or hallucinated content 21 00:00:45.570 --> 00:00:48.870 and how we can tell ChatGPT to avoid hallucinating. 22 00:00:48.870 --> 00:00:50.587 So you can see here we've had a prompt, 23 00:00:50.587 --> 00:00:53.700 "Who was the first man to visit to Jupyter?" 24 00:00:53.700 --> 00:00:55.170 And I've just said, "Make something up." 25 00:00:55.170 --> 00:00:57.360 This is obviously not factual, it's not correct. 26 00:00:57.360 --> 00:00:59.617 Now obviously, I've had to say something like, 27 00:00:59.617 --> 00:01:00.630 "Make something up," 28 00:01:00.630 --> 00:01:03.750 but there will be scenarios where you ask ChatGPT something 29 00:01:03.750 --> 00:01:05.040 and it will come up with something 30 00:01:05.040 --> 00:01:07.080 that's not technically true. 31 00:01:07.080 --> 00:01:08.250 How can we get over that? 32 00:01:08.250 --> 00:01:11.580 How can we ground ChatGPT in some knowledge 33 00:01:11.580 --> 00:01:13.320 that either is at your own company 34 00:01:13.320 --> 00:01:14.640 or your personal information? 35 00:01:14.640 --> 00:01:17.310 So I'm gonna give you a technique that you can use for this. 36 00:01:17.310 --> 00:01:20.740 So I'm gonna go to Google and type digital marketing 37 00:01:22.200 --> 00:01:23.940 and I'm gonna take some information here. 38 00:01:23.940 --> 00:01:25.080 And this could be, you know, 39 00:01:25.080 --> 00:01:27.420 you might be using information from your own company 40 00:01:27.420 --> 00:01:29.160 or it could be a PDF. 41 00:01:29.160 --> 00:01:30.780 I'm gonna create a new chat. 42 00:01:30.780 --> 00:01:33.210 I'm gonna triple quote 43 00:01:33.210 --> 00:01:36.450 and then I'm gonna paste this information in like this. 44 00:01:36.450 --> 00:01:37.717 And I'm gonna say, 45 00:01:37.717 --> 00:01:40.030 "I want you to answer 46 00:01:41.130 --> 00:01:43.410 the following query 47 00:01:43.410 --> 00:01:48.180 by only referencing the documents above 48 00:01:48.180 --> 00:01:51.960 in triple quote characters. 49 00:01:51.960 --> 00:01:56.190 If you don't have an answer, 50 00:01:56.190 --> 00:01:58.143 simply respond with, 51 00:01:59.190 --> 00:02:00.507 I don't know." 52 00:02:01.440 --> 00:02:06.210 And then we can ask ChatGPT, what is the weather, right? 53 00:02:06.210 --> 00:02:09.150 And basically you'll see that we've now grounded ChatGPT 54 00:02:09.150 --> 00:02:11.880 in some knowledge right from the internet. 55 00:02:11.880 --> 00:02:16.880 So we can also say, "What is social media marketing?" 56 00:02:17.490 --> 00:02:19.500 And what you'll see is that the knowledge 57 00:02:19.500 --> 00:02:21.540 that was contained within those triple quotes 58 00:02:21.540 --> 00:02:23.790 was actually factually grounded 59 00:02:23.790 --> 00:02:26.910 based off some information that I chose to ground it in. 60 00:02:26.910 --> 00:02:28.080 So this is a way where 61 00:02:28.080 --> 00:02:29.970 if you've got some company information 62 00:02:29.970 --> 00:02:32.850 or you have some specific documents 63 00:02:32.850 --> 00:02:34.380 or certain types of knowledge, 64 00:02:34.380 --> 00:02:36.660 you could put that inside triple quotes 65 00:02:36.660 --> 00:02:37.987 and you can say to ChatGPT, 66 00:02:37.987 --> 00:02:40.500 "I only want you to answer from this knowledge." 67 00:02:40.500 --> 00:02:42.270 And then you're grounding ChatGPT 68 00:02:42.270 --> 00:02:44.370 in not knowledge that it knows internally, 69 00:02:44.370 --> 00:02:46.410 but knowledge that you've decided is important 70 00:02:46.410 --> 00:02:47.850 for answering your question. 71 00:02:47.850 --> 00:02:50.550 So this is one common use case 72 00:02:50.550 --> 00:02:54.060 and technique that you can use to overcome hallucinations 73 00:02:54.060 --> 00:02:55.980 for when factual consistency 74 00:02:55.980 --> 00:02:57.990 is really important in the answers. 75 00:02:57.990 --> 00:03:00.510 Another way that you can factually ground a reference 76 00:03:00.510 --> 00:03:04.260 inside of ChatGPT is adding a file to a conversation. 77 00:03:04.260 --> 00:03:06.390 For example, we can go and hit the plus button here 78 00:03:06.390 --> 00:03:08.340 and then go to add photos and files. 79 00:03:08.340 --> 00:03:11.250 I'm gonna put a file that you can download in this lecture 80 00:03:11.250 --> 00:03:12.480 to follow along with. 81 00:03:12.480 --> 00:03:15.180 Go and download a digital company profile. 82 00:03:15.180 --> 00:03:16.777 And then I'm gonna say, 83 00:03:16.777 --> 00:03:21.537 "What is the company Aurora?" 84 00:03:22.380 --> 00:03:25.770 And because you've added a file, you can ground ChatGPT 85 00:03:25.770 --> 00:03:28.410 to answer specifically from that documentation. 86 00:03:28.410 --> 00:03:29.940 You can add lots of different files 87 00:03:29.940 --> 00:03:32.700 and also images directly inside of ChatGPT 88 00:03:32.700 --> 00:03:34.470 to help sure that it doesn't hallucinate. 89 00:03:34.470 --> 00:03:36.300 In the next video, we're gonna have a look at the difference 90 00:03:36.300 --> 00:03:38.670 between two types of models that have come out recently. 91 00:03:38.670 --> 00:03:42.030 So there is chat models, which are the older type of models, 92 00:03:42.030 --> 00:03:45.150 the newer type of models, which are reasoning models. 93 00:03:45.150 --> 00:03:47.250 And we'll look at the differences between those two 94 00:03:47.250 --> 00:03:48.750 and how you can get started with them. 95 00:03:48.750 --> 00:03:50.313 Cool, see you in the next one.