WEBVTT 00:00.140 --> 00:05.660 Large language models with 100 billion parameters can achieve great things. 00:05.660 --> 00:14.000 They can output us strong results from a wide variety and range of tasks, and they can do it even with 00:14.000 --> 00:17.090 no training or little training by us. 00:17.150 --> 00:25.100 However, even the largest LLM model can still struggle with certain kinds of problems. 00:25.100 --> 00:33.680 Those multi-step reasoning tasks, such as math word problems and common sense reasoning that are easy 00:33.680 --> 00:37.400 for us humans are not so easy for the models. 00:37.400 --> 00:44.150 So for that, researchers at Google came up with a method which is called chain of thought and chain 00:44.180 --> 00:44.480 of thought. 00:44.480 --> 00:52.070 Prompting aims to improve the reasoning abilities of the llms, and this method enables the models to 00:52.100 --> 00:58.970 decompose multi-step problems into intermediate steps, allowing the model to solve complex reasoning 00:58.970 --> 01:07.820 problems that are not solvable with standard prompting methods or Chain of thought, or abbreviated 01:07.850 --> 01:15.740 Cot, is breaking down a problem into a series of intermediate reasoning steps, and it significantly 01:15.740 --> 01:19.340 improved the ability of llms to perform complex reasoning. 01:19.490 --> 01:26.600 So by breaking down a complex problem into smaller, more manageable steps, the chain of thought allows 01:26.690 --> 01:31.160 llms to reason more accurate and to have better efficiency. 01:31.580 --> 01:38.660 So, like most prompt engineering techniques, the chain of thought technique was first introduced in 01:38.660 --> 01:41.510 an official paper for prompt Engineering. 01:41.510 --> 01:46.850 You can see the paper right over here and its URL also listed in the course's resources. 01:48.050 --> 01:55.280 Okay, so let's start introducing this topic from the examples that are noted in this paper over here. 01:55.310 --> 02:01.910 Okay, so we want to start with the standard problem prompting and introduce the problem and to see 02:01.910 --> 02:03.200 that it's not sufficient. 02:03.230 --> 02:03.590 Okay. 02:03.620 --> 02:08.300 And we'll see how the chain of thought prompting resolves this insufficiency. 02:08.630 --> 02:13.550 So we have two prompts over here that we asked our AI model. 02:13.580 --> 02:16.580 The first one is a question and it's kind of a riddle. 02:16.580 --> 02:19.460 So Sean has five toys for Christmas. 02:19.460 --> 02:23.000 He got two toys each from his mom and his dad. 02:23.000 --> 02:25.490 And how many toys does he has in total? 02:25.490 --> 02:30.650 So the AA model A responded with nine, which is the correct answer. 02:30.680 --> 02:33.380 Okay, so basically he had five toys already. 02:33.410 --> 02:36.020 Two he got from his mom and two he got from his dad. 02:36.020 --> 02:38.900 So it's five plus two plus two. 02:38.930 --> 02:39.980 This is great. 02:40.250 --> 02:46.970 Now afterwards those researchers, they asked another riddle which is kind of similar. 02:47.150 --> 02:48.740 It has the same thought process. 02:48.740 --> 02:53.270 So the question is the prompt is John takes care of ten dogs. 02:53.270 --> 02:58.250 Each dog takes five hours a day to walk and take care of their business. 02:58.250 --> 03:03.170 How many hours a week does John need to take care of his dogs? 03:03.170 --> 03:06.140 So the AI model outputted. 03:06.140 --> 03:12.470 That was 50, so probably the calculation was ten times five, which outputted 50. 03:12.500 --> 03:14.870 Now this is not the correct solution. 03:14.900 --> 03:18.020 The correct answer for this is 35 hours. 03:18.050 --> 03:24.080 So in the standard prompting that we saw right now, which is actually if we classify it, it goes to 03:24.110 --> 03:27.770 the topic of zero shot prompting. 03:27.770 --> 03:31.160 Then we got insufficient answers for our questions. 03:31.190 --> 03:31.490 Okay. 03:31.520 --> 03:33.950 So we had kind of a limitation over here. 03:33.950 --> 03:38.540 So the paper here introduced this limitation with the standard prompting. 03:38.540 --> 03:41.780 And here was an example of a zero shot prompting. 03:41.930 --> 03:47.990 And the paper suggested that there is a new way of prompting technique which is called change of thought. 03:48.050 --> 03:51.110 So chain of thought is actually pretty simple. 03:51.110 --> 03:58.580 What we want to do is guide the model to solve the problem, like we as humans solve those problems. 03:58.580 --> 04:05.030 And by doing that, we'll make the model smarter and better, and hopefully it will output us the answer 04:05.030 --> 04:07.160 that we want and the correct answer. 04:07.190 --> 04:07.640 Okay. 04:07.670 --> 04:14.960 So in the example over here, what those researchers did, they took the first question with the toys 04:14.960 --> 04:17.430 and they gave an example answer. 04:17.460 --> 04:21.690 Okay, so you can think of it as a one shot prompting. 04:21.720 --> 04:22.290 Okay. 04:22.290 --> 04:24.390 But they did more than that. 04:24.390 --> 04:29.850 They also supplied the chain of thought how we as humans solve this problem. 04:29.880 --> 04:33.900 So in their answer, they said that John started with five toys. 04:33.900 --> 04:36.900 He got two from his mom and two from his dad. 04:36.900 --> 04:42.120 So the calculation should be four plus two plus two, which is nine. 04:42.150 --> 04:45.330 Okay, that's why the answer is nine. 04:45.360 --> 04:45.870 Okay. 04:45.900 --> 04:49.260 And then they asked the same question again. 04:49.290 --> 04:55.590 And it turned out that the model was smart enough to take the chain of thought that the researchers 04:55.590 --> 05:01.590 applied to it for the previous question, which was kind of similar, and to apply this chain of thought 05:01.590 --> 05:03.030 in the next answer. 05:03.060 --> 05:03.540 Okay. 05:03.570 --> 05:08.700 So you can see that the second response with the chain of thought prompting. 05:08.700 --> 05:14.640 Then the model got the correct answer and it actually supplied the steps that it made in order to get 05:14.640 --> 05:15.390 to that answer. 05:15.420 --> 05:20.700 Okay, so the AI model outputted that John took care of ten dogs. 05:20.700 --> 05:25.080 And because each dog takes half an hour a day to walk and take care of. 05:25.110 --> 05:30.630 So the calculation should be ten, which is ten dogs times the time it takes to take care of them, 05:30.630 --> 05:32.670 which is five hours per day. 05:32.700 --> 05:39.720 So if we have five hours per day for taking care of the dogs and we have seven days in a week, then 05:39.720 --> 05:44.610 the calculation should be five times seven, which is going to be equal to 35. 05:44.640 --> 05:45.060 Okay. 05:45.090 --> 05:50.700 So the thought process was pretty much the same over here with both questions. 05:50.700 --> 05:57.600 It was to take the calculation that we need to calculate and to break it down to two sub calculation. 05:57.630 --> 05:58.110 Okay. 05:58.140 --> 05:59.190 So yeah. 05:59.190 --> 06:04.770 So that was the chain of thought prompting in the research paper that you see below. 06:05.940 --> 06:13.290 Chain of thought is a significant development in prompt engineering and LM usages because it allows 06:13.290 --> 06:20.340 the LM model to approach problem solving in a more human like manner, and it allows the models to break 06:20.340 --> 06:26.430 down complex problems into intermediate steps that are resolved individually. 06:26.460 --> 06:34.890 So this is a major step in prompt engineering which opens us to solving a lot more problems. 06:35.760 --> 06:41.970 So we saw in earlier videos that we had zero shot, one shot and few shot prompting. 06:41.970 --> 06:46.200 So in chain of thought we actually have something very similar. 06:46.230 --> 06:46.650 Okay. 06:46.680 --> 06:51.390 So we have something which is called a zero shot chain of thought. 06:51.390 --> 07:01.440 So in this kind of prompt we simply add to our prompt something like let's think step by step and then 07:01.440 --> 07:07.500 the model will output the answer while explaining what it gets to be the steps. 07:07.530 --> 07:14.790 Okay, so we did not supply the reasoning engine to the model, and we let it have its creative freedom 07:14.790 --> 07:15.690 and choose for us. 07:15.720 --> 07:16.320 Okay. 07:16.320 --> 07:24.300 And if we integrate few shots chain of thought prompting, then we simply in our prompt supply the model 07:24.300 --> 07:30.300 with the answer we want and explain to them what are the steps and what are the chain of thought that 07:30.300 --> 07:32.520 we used to resolve those steps? 07:32.520 --> 07:34.800 And how did we come up with this answer. 07:34.800 --> 07:40.080 And the model takes this data, processes it, and then outputs the correct answer. 07:40.110 --> 07:40.650 Okay. 07:40.680 --> 07:44.520 So zero shot chain of thought can be actually beneficial. 07:44.550 --> 07:44.880 Okay. 07:44.910 --> 07:48.660 And you can actually see the reasoning process of the AI model. 07:48.690 --> 07:49.080 Okay. 07:49.110 --> 07:51.660 But it will have no prior knowledge. 07:51.660 --> 07:59.370 And in the few shot change of thought prompt we will attach to our prompt the real answer that we're 07:59.370 --> 08:06.600 expecting in a very similar problem and explain how we got to this answer and the AI model will take 08:06.600 --> 08:14.280 it, will process it, will learn it, and then the model will be able to deduce the chain of thought 08:14.310 --> 08:21.480 that we supplied it to other problems, and will be able to apply this chain of thought and will be 08:21.510 --> 08:28.920 able to solve similar, but not really the same problems in the way that we wanted it to resolve it.