WEBVTT 00:00.000 --> 00:02.220 Eden: Let's discuss what's flow engineering, 00:02.220 --> 00:06.240 which is a new idea that has recently been discussed 00:06.240 --> 00:08.760 in the generative AI community. 00:08.760 --> 00:10.380 Now I warn you in advance, 00:10.380 --> 00:12.720 this is a very theoretical video 00:12.720 --> 00:16.260 and the idea of flow engineering is pretty abstract 00:16.260 --> 00:19.530 and not yet formalized to the end. 00:19.530 --> 00:20.880 So I don't expect you 00:20.880 --> 00:23.220 to understand everything in this video, 00:23.220 --> 00:25.380 but I promise you, by the end of this course, 00:25.380 --> 00:28.143 you'll know exactly what flow engineering is all about. 00:30.030 --> 00:33.660 Flow engineering is a systematic and strategic approach 00:33.660 --> 00:37.140 for developing software that will incorporate 00:37.140 --> 00:39.720 AI-driven decision-making processes. 00:39.720 --> 00:42.390 The essential goal of flow engineering 00:42.390 --> 00:44.100 is to manage and optimize 00:44.100 --> 00:47.940 how AI systems with LLMs handle tasks 00:47.940 --> 00:52.830 by defining a clear flow or a sequence of operations. 00:52.830 --> 00:54.870 These flows are not merely linear, 00:54.870 --> 00:58.230 so they may involve complex decision-making nodes 00:58.230 --> 01:01.170 where the AI may generate multiple outputs, 01:01.170 --> 01:05.220 which are often assessed and refined in an iterative cycle. 01:05.220 --> 01:08.190 Flow engineering is a structured process 01:08.190 --> 01:12.240 in AI development systems which guides the AI 01:12.240 --> 01:14.760 through a series of well-defined steps 01:14.760 --> 01:18.093 in order to improve the output quality of the AI system. 01:19.890 --> 01:23.310 Flow engineering aims to incorporate systematic planning 01:23.310 --> 01:24.750 and testing phase 01:24.750 --> 01:27.690 that mimic human development processes, 01:27.690 --> 01:30.570 and all of this to enhance reliability 01:30.570 --> 01:34.143 and functionality of AI-generated solutions. 01:37.140 --> 01:40.140 One of the issues with AutoGPT 01:40.140 --> 01:44.430 and other autonomous agents projects like BabyAGI, 01:44.430 --> 01:46.320 they are long-term planning. 01:46.320 --> 01:50.250 So they receive a goal, something to achieve, 01:50.250 --> 01:53.160 and then they start breaking it down to tasks, 01:53.160 --> 01:57.090 and then they start implementing and executing those tasks 01:57.090 --> 02:01.680 and derive subtasks of those tasks and so on and so on. 02:01.680 --> 02:05.010 And de facto, they aren't really working. 02:05.010 --> 02:08.550 So the point here is that we as developers 02:08.550 --> 02:11.550 want to give the instructions of what to do. 02:11.550 --> 02:14.610 We don't want the AI to start going and creating 02:14.610 --> 02:17.700 some imaginary attempts and start executing them. 02:17.700 --> 02:20.370 This is highly prone for problems 02:20.370 --> 02:23.880 and for the LLM to simply go out of hand. 02:23.880 --> 02:27.960 We as developers want to tell the LLM exactly what to do. 02:27.960 --> 02:30.360 We want to define the tasks and we want the LLM 02:30.360 --> 02:33.093 to stay within the context of those tasks. 02:34.020 --> 02:36.690 There are still decisions the LLM could make. 02:36.690 --> 02:39.240 For example, determining if the output 02:39.240 --> 02:40.920 is ready to be released 02:40.920 --> 02:43.020 and if it's a good enough solution 02:43.020 --> 02:45.570 or which step now to take. 02:45.570 --> 02:47.460 However, we, as developers, 02:47.460 --> 02:50.610 we define the scope and we do most of the planning 02:50.610 --> 02:51.443 for the LLM 02:51.443 --> 02:55.770 and the LLM can work within the flow that we have created. 02:55.770 --> 02:59.160 We give the LLM the blueprints to follow. 02:59.160 --> 03:01.890 And if we think about it as a state machine, 03:01.890 --> 03:04.260 we, as developers, we write the states. 03:04.260 --> 03:07.560 So we write what is the flow and what's need to be done, 03:07.560 --> 03:09.150 and what steps do we have. 03:09.150 --> 03:12.570 However, we can incorporate an LLM to decide 03:12.570 --> 03:15.540 which flow to take based on the input it gets. 03:15.540 --> 03:17.670 For example, whether to curate an answer 03:17.670 --> 03:20.160 or to simply release it to the user. 03:20.160 --> 03:21.420 And in the state machine 03:21.420 --> 03:23.400 is whether to go to step i, 03:23.400 --> 03:26.160 or maybe to i plus two, or whatever. 03:26.160 --> 03:28.740 So the state machine is an engineering decision, 03:28.740 --> 03:30.963 and no statistics is involved there. 03:33.240 --> 03:36.360 And LangGraph in the context of flow engineering 03:36.360 --> 03:38.790 is implementing this middle ground. 03:38.790 --> 03:41.910 It's the intermediate between fully autonomous agents 03:41.910 --> 03:44.460 that can decide what to do and how to do it 03:44.460 --> 03:47.790 and the LangChain chain, which is fully deterministic 03:47.790 --> 03:51.600 and there is no flexibility within this flow. 03:51.600 --> 03:53.400 So with flow engineering, 03:53.400 --> 03:57.390 we can achieve complex agentic solutions 03:57.390 --> 04:00.030 which will require us to build a state machine 04:00.030 --> 04:02.730 to define the steps of our flow. 04:02.730 --> 04:06.600 And we can use an LLM to either be a part of the step, 04:06.600 --> 04:10.650 for example, to make an LLM call to generate a tweet, 04:10.650 --> 04:13.110 and it can also be something 04:13.110 --> 04:15.570 that will tell us which steps to go to. 04:15.570 --> 04:19.530 So we leverage the LLM to do stuff within a step 04:19.530 --> 04:23.370 and we can leverage the LLM to decide which step to go. 04:23.370 --> 04:26.040 But the point here is that we, as developers, 04:26.040 --> 04:29.163 we define the flow and we have full control of it. 04:31.200 --> 04:34.320 And in LangGraph we can define graphs, 04:34.320 --> 04:36.540 and those are going to be nodes and edges, 04:36.540 --> 04:39.960 and we can have cycles in them, and we can build, 04:39.960 --> 04:41.280 and we'll do this in the course, 04:41.280 --> 04:43.500 very advanced logic 04:43.500 --> 04:47.580 that will give us a very complex AI system. 04:47.580 --> 04:50.490 And this is an example of a graph 04:50.490 --> 04:53.070 that is going to write a tweet, 04:53.070 --> 04:56.400 but it's going to also reflect and critique this tweet 04:56.400 --> 04:57.990 and do this in iterations 04:57.990 --> 05:00.663 until we get a very nice Twitter post. 05:03.060 --> 05:06.810 I believe that in the near future when we'll develop 05:06.810 --> 05:09.750 AI software and generative AI software, 05:09.750 --> 05:11.670 then the time we'll spend on the software 05:11.670 --> 05:13.470 will distribute the following. 05:13.470 --> 05:15.840 60% is going to go to flow engineering 05:15.840 --> 05:17.370 and into architecture, 05:17.370 --> 05:21.240 the state machine, the nodes of what steps can we make, 05:21.240 --> 05:23.760 maybe we can incorporate an LLM within them 05:23.760 --> 05:27.090 and maybe even the LLM can decide which step to do. 05:27.090 --> 05:29.940 However, 35% will go to fine-tuning, 05:29.940 --> 05:33.000 so to make the model very specific for the tasks 05:33.000 --> 05:34.080 that we need to achieve. 05:34.080 --> 05:36.663 And 5% will be dedicated to prompt engineering. 05:39.150 --> 05:39.983 Alrighty. 05:39.983 --> 05:43.380 So I know that this video was super abstract 05:43.380 --> 05:46.170 and you probably didn't understand everything. 05:46.170 --> 05:47.100 Don't worry. 05:47.100 --> 05:50.550 I know this is a new topic and it's kind of ambiguous. 05:50.550 --> 05:53.580 However, I promise you that by the end of this course, 05:53.580 --> 05:56.520 you'll know exactly what flow engineering is 05:56.520 --> 05:58.860 and you'll understand why it is so important 05:58.860 --> 06:01.260 when we build those advanced agents 06:01.260 --> 06:02.960 using flow engineering techniques.