WEBVTT 00:00.080 --> 00:00.840 There you have it. 00:00.840 --> 00:06.280 Those are the five tricks that allow us to take calls to LMS with inputs and the way we interpret the 00:06.280 --> 00:06.840 outputs. 00:07.080 --> 00:14.160 Do it in such a way that gives this impression that we have something out there carrying out tasks autonomously 00:14.160 --> 00:14.880 for us. 00:15.040 --> 00:19.160 And don't worry if you if you didn't get all of this because we're going to be doing this again and 00:19.160 --> 00:23.840 again and again, and by the end of it it's all going to connect wonderfully, but hopefully this giving 00:23.840 --> 00:25.680 you some intuition for it. 00:25.800 --> 00:26.920 So what about the trap? 00:26.920 --> 00:28.920 I told you there were five tricks and a trap. 00:29.000 --> 00:34.120 So this trap is something that I find personally galling, and it happens all the time. 00:34.120 --> 00:38.520 And so I wanted to to warn you about it and get your take on it. 00:38.560 --> 00:40.880 And I'm calling it the human trap. 00:41.120 --> 00:45.600 The proper word people use for this is anthropomorphizing. 00:45.880 --> 00:53.840 It's the problem of when people approach a genetic AI, treating llms as if they're humans, as if they 00:53.840 --> 00:55.760 have roles and responsibilities. 00:55.760 --> 00:57.960 And let me tell you more about what I mean. 00:58.160 --> 01:03.710 It's so common when I hear business people have have an objective they want to automate a business process, 01:03.710 --> 01:05.830 something we'll be doing a lot on this course. 01:05.990 --> 01:11.990 And the first their go to thought for business people and for technology people too, for engineers 01:11.990 --> 01:19.750 that should know better, uh, is to say, okay, let's create an agent architecture, an agent architecture, 01:19.750 --> 01:23.830 which is a diagram with different agents with lines between them. 01:24.030 --> 01:30.230 And I'm going to assign roles and responsibilities to these agents just by sort of analogy with the 01:30.230 --> 01:33.590 way humans go about doing this as if they are people. 01:33.590 --> 01:39.870 I'm going to have agents that represent different jobs, and I'm guilty of doing this myself. 01:39.870 --> 01:44.430 I do this in some of my, my, my courses teaching a genetic engineering, and we build things like 01:44.430 --> 01:47.350 a trading floor and we have traders and researchers. 01:47.350 --> 01:50.070 It's so natural to go towards that. 01:50.070 --> 01:53.630 And when you're doing like, like toy projects and demos, it's fine. 01:53.630 --> 01:59.310 You can do that because it's great fun to see it in progress, but it's not a disciplined way to do 01:59.310 --> 01:59.670 it. 01:59.710 --> 02:02.350 And it has lots of lots of problems. 02:02.350 --> 02:03.260 So what are the problems? 02:03.300 --> 02:10.900 Well, the big problem is that you have to keep in mind that what llms are good at doing is generating 02:10.900 --> 02:12.820 realistic content. 02:12.820 --> 02:17.220 That's what they're trained to do, generating stuff that seems compelling. 02:17.380 --> 02:22.260 And so if you give it a job, you say you are an evaluation agent. 02:22.260 --> 02:24.940 Your job is to evaluate what came before. 02:25.100 --> 02:26.900 You should give it marks out of ten. 02:27.140 --> 02:29.780 Then it will do that because that's what it's meant to do. 02:29.820 --> 02:33.660 And it will come up with a reason, because you've told it to give a reason. 02:33.860 --> 02:36.300 It doesn't mean that it's doing it well. 02:36.300 --> 02:40.380 It doesn't mean that this evaluation is aligned with your objectives. 02:40.500 --> 02:43.060 It just means it's going to follow the script. 02:43.060 --> 02:45.580 It's going to do what it's told to do in the prompt. 02:45.740 --> 02:50.780 So the danger, the risk of this is that you can fool yourself into thinking you have this, this, 02:50.780 --> 02:55.940 this whole sort of cadre, this whole group of different agents all doing their thing, all generating 02:55.940 --> 02:58.860 very reasonable, very realistic content. 02:58.860 --> 03:01.580 But it might all be llms slop. 03:01.620 --> 03:04.370 It might be nonsense that apparently is doing something. 03:04.370 --> 03:06.650 They're all they're all collaborating together. 03:06.650 --> 03:11.970 They're all assuming their roles, but they're not actually solving your task in an accurate way. 03:12.010 --> 03:16.490 So it's all very well, me complaining, but what am I suggesting is the the answer? 03:16.730 --> 03:18.530 Well, the right way to do it. 03:18.570 --> 03:25.290 When you divide up your problem into different agents, you should do it because it solves your problem 03:25.290 --> 03:30.770 better, not because it sounds like those are the roles you would have, but because you've tried it 03:30.770 --> 03:32.370 and that's an improvement. 03:32.370 --> 03:34.290 You've got some chunky problem. 03:34.290 --> 03:37.010 It makes sense to divide it into some steps. 03:37.010 --> 03:40.850 You try it and you get better outcomes and then you stick with it. 03:40.850 --> 03:42.210 That's the way to do it. 03:42.330 --> 03:48.170 Uh, scientifically, with experiments and most importantly with this, the magical word is that word 03:48.170 --> 03:49.410 in that last bullet there. 03:49.610 --> 03:50.650 Evaluate. 03:50.650 --> 03:56.890 You need to have a way to measure your outcomes, and you should divide things into smaller steps or 03:56.890 --> 04:03.410 organize your agents differently because you get better evaluations, because it results in superior 04:03.450 --> 04:04.210 performance. 04:04.320 --> 04:10.040 That's the right way to do it, not just because it sounds like that's the right kind of responsibility. 04:10.080 --> 04:15.880 Now, often it's a good starting point to start by analogy with with some human organization. 04:15.880 --> 04:20.200 If we divided up roles and responsibilities for for humans, we probably had some, some reason for 04:20.240 --> 04:20.640 that. 04:20.640 --> 04:26.520 Maybe it's a decent starting point, but you should treat it as your starting point and always start 04:26.520 --> 04:28.240 as simple as possible. 04:28.360 --> 04:32.920 Start with just one role, divide up and start assigning more responsibilities. 04:32.920 --> 04:38.640 Experiment with that because it's going to help you get better outcomes and then measure it. 04:38.640 --> 04:40.000 That's the right way to do it. 04:40.000 --> 04:41.720 That's how to avoid the trap. 04:41.720 --> 04:45.200 And that's a wrap on the theory around Agentic AI. 04:45.240 --> 04:51.840 I hope this has given you some good intuition as we dive into building agentic workflows and also some 04:51.840 --> 04:57.560 some real, real world kind of lessons learned from actually building Agentic AI in anger. 04:57.960 --> 05:03.320 And now it's time for us to go back to N810, starting with the navigation. 05:03.320 --> 05:06.240 The big picture building blocks of N810.