WEBVTT 00:00.040 --> 00:01.400 Well, this is good news. 00:01.400 --> 00:03.440 You've you've decided to come back for more. 00:03.480 --> 00:06.280 You weren't put off thoroughly by what we did yesterday. 00:06.320 --> 00:07.600 You're back for more. 00:07.640 --> 00:08.520 Welcome. 00:08.520 --> 00:10.000 Welcome to week one. 00:10.040 --> 00:10.720 Day two. 00:10.880 --> 00:13.920 Now, look, I am a a learning by doing type. 00:13.920 --> 00:18.760 The way that that I like to teach is the way that I like to learn myself, which is by going in there 00:18.760 --> 00:20.640 into a product and using it. 00:21.080 --> 00:23.200 But today is a little bit of an odd one out. 00:23.200 --> 00:26.200 Today it's going to be more talking than usual. 00:26.240 --> 00:31.200 I'm going to be giving you some foundational, some some basis for everything that is to come. 00:31.200 --> 00:36.440 We're going to be talking about Agentic, AI and also some of the constructs around ten, but there'll 00:36.440 --> 00:38.400 be a bit of doing as well, never fear. 00:38.640 --> 00:39.840 Let's get started. 00:39.840 --> 00:42.560 Let me tell you more about Agentic AI. 00:42.880 --> 00:47.200 Okay, so diving right in the first topic is AI agents. 00:47.200 --> 00:49.400 Before we go back to Na ten again. 00:49.760 --> 00:52.200 And this is a quick recap of what we already said yesterday. 00:52.200 --> 00:55.240 There are so many ways to define an AI agent. 00:55.240 --> 00:59.360 It's almost become a joke, but some of the the ways that are most common. 00:59.360 --> 01:06.730 Originally, companies like OpenAI described it as systems that can work for you independently, like 01:06.770 --> 01:11.010 the operator agent now GPT ChatGPT agent. 01:11.170 --> 01:18.290 And then there came this this emerging definition, which was about AI systems, where an LLM decides 01:18.290 --> 01:19.570 what to do next. 01:19.610 --> 01:25.170 It generates content and that content describes, hey, let's do A and then B, and then C, or maybe 01:25.210 --> 01:31.370 we should do A and then C and then b, and that, that, that is the output that comes from calling 01:31.370 --> 01:32.170 an LLM. 01:32.170 --> 01:34.930 And we interpret that output and say okay that's what we'll do. 01:35.010 --> 01:41.410 So that idea, that concept where an LLM decides the order of activities, orchestrates the workflow 01:41.530 --> 01:43.650 that became the working definition. 01:43.650 --> 01:50.210 But more recently it has become this idea that that an agent is where you have an LLM running tools 01:50.450 --> 01:52.450 in a loop to achieve a goal. 01:52.730 --> 01:56.570 Okay, let's let's go deeper into agent K. 01:56.730 --> 02:00.890 There's a lot about Agentic AI that feels kind of magical. 02:01.330 --> 02:08.470 But sitting behind all of that, the reality of Agentic AI is actually something that's quite pedestrian. 02:08.670 --> 02:16.230 Most of the magical stuff in AI just comes down to writing good prompts to llms. 02:16.710 --> 02:19.990 The prompt, of course, is just what we call the input. 02:20.030 --> 02:24.270 The text that's sent to an LLM and LLM, a large language model. 02:24.270 --> 02:31.390 As we said, it's this this data science piece of of this algorithm, this bit of statistics that takes 02:31.390 --> 02:34.030 an input and it generates an output. 02:34.030 --> 02:39.790 And the way that it, that it does that is by trying to figure out what's the most likely thing to come 02:39.790 --> 02:44.350 next after this input and this input, we call the prompt. 02:44.550 --> 02:49.390 And when you actually call one of these llms, every call is stateless. 02:49.430 --> 02:51.750 It takes that input, it generates the output. 02:51.750 --> 02:53.630 It doesn't know how it was called before. 02:53.630 --> 02:56.310 It's been called by all sorts of people around the world all the time. 02:56.310 --> 03:01.950 If it's like GPT, it gets an input sequence and it generates an output sequence based on it. 03:01.950 --> 03:05.070 And what we can do with two things that we can do. 03:05.110 --> 03:10.360 We can decide what input to send it, and we can decide what we're going to do with the output. 03:10.360 --> 03:16.040 And most of our genetic AI is about those two activities getting the right input and interpreting the 03:16.040 --> 03:16.640 output. 03:16.680 --> 03:18.120 And that's really it. 03:18.160 --> 03:23.280 Now, I imagine many of you know this already, but typically this input sequence, this input text 03:23.280 --> 03:26.440 that goes to an LLM consists of a number of different parts. 03:26.680 --> 03:30.320 It often starts with something called the system prompt. 03:30.360 --> 03:34.840 And the system prompt is a bit of text that gives the overall context. 03:34.880 --> 03:36.120 It sets the scene. 03:36.120 --> 03:41.800 It tells the LLM what role it's playing, any background information, any style that it should respond 03:41.800 --> 03:42.200 in. 03:42.240 --> 03:44.720 That's that's in the part that's called the system prompt. 03:44.720 --> 03:47.920 And then often after that comes a user prompt. 03:47.920 --> 03:53.920 And that is the specific message which the LLM should respond to in the content it generates. 03:54.160 --> 03:57.880 And you can also have in there the whole conversation so far. 03:57.920 --> 04:00.480 The original user prompt that was like hi there. 04:00.920 --> 04:07.040 The llms response to that, the next user prompt, another response, and finally the user prompt that 04:07.040 --> 04:08.360 it's now responding to. 04:08.520 --> 04:11.650 All of that can go in this one input sequence. 04:11.970 --> 04:18.050 And this works well because when we trained this LLM, when it was given tons of data to analyze so 04:18.050 --> 04:23.050 that it can match patterns in the future, that data was given to train was organized this way. 04:23.090 --> 04:27.890 Typically with this system prompt at the start and then user and an assistant response, another user 04:27.890 --> 04:28.490 message. 04:28.490 --> 04:36.010 And because it's seen lots of data in that format, it's good at generating coherent outputs that sort 04:36.010 --> 04:38.090 of adheres to this structure. 04:38.090 --> 04:41.850 And that's why it works so well, because the training data was organized that way. 04:42.210 --> 04:51.090 And I mentioned this before, but the the key is that we get to decide how to organize that input data, 04:51.090 --> 04:58.010 and we can do all sorts of tricks to it to make it most likely that what the LLM generates, the outputs 04:58.010 --> 05:00.930 that it's pattern matching decides are the right outputs. 05:01.130 --> 05:07.090 It's going to be consistent with our business objective, so we can come up with clever techniques that 05:07.090 --> 05:11.730 we will go through to make it more and more likely that we'll be able to get the right kind of output. 05:12.010 --> 05:17.740 And when I say the right kind of output, we can choose how we interpret that output, and we can interpret 05:17.740 --> 05:23.780 it in different ways at our whim so that it can achieve something, something cool that we want to do. 05:24.140 --> 05:28.580 As I say, I gave the example before that the input could say something like please respond with what 05:28.620 --> 05:30.020 order I should do things. 05:30.020 --> 05:34.620 And then the output might be you should do things in this order, and we can interpret that and use 05:34.620 --> 05:37.980 that to then execute other steps in a process. 05:37.980 --> 05:42.900 In other words, we're treating this text output as if it's controlling a workflow. 05:43.260 --> 05:49.980 So this idea of deciding how we will use this input and how we will interpret the output, this this 05:49.980 --> 05:53.900 is really the sort of secret sauce that makes up a genetic AI. 05:54.100 --> 05:56.700 It all comes down to clever prompting. 05:56.700 --> 05:59.660 And sometimes people have have changed the language a bit. 05:59.700 --> 06:02.820 In the old days we used to talk about prompt engineering, which is a bit old hat. 06:02.860 --> 06:06.500 Now, the idea of figuring out the best way to prompt an LLM. 06:06.700 --> 06:13.340 The new term is context engineering, which is about how best to set up this context in what gets sent 06:13.340 --> 06:18.200 to the LLM to really position it to achieve our business objective. 06:18.240 --> 06:19.480 That's what it's all about. 06:19.800 --> 06:26.680 So most of the magic behind the sky can be explained with five tricks. 06:26.800 --> 06:29.480 Five conjuring tricks that we use. 06:29.680 --> 06:35.280 That's all built around how we prompt and a little bit of how we train LMS. 06:35.560 --> 06:42.480 And then I also want to mention a trap, which is a common pitfall, a common mistake that people can 06:42.480 --> 06:43.880 make with their genetic AI. 06:44.080 --> 06:49.640 And so the five tricks that we will cover there is something called the illusion of memory, which is 06:49.640 --> 06:52.280 a fairly simple one that probably many of you know already. 06:52.480 --> 06:56.800 There's the thinking and reasoning idea and reasoning budgets. 06:57.080 --> 07:02.720 There's chaining LMS together, there's the use of tools, what it is and what it's not. 07:02.720 --> 07:07.600 And then there is the famous agent loop and then the trap. 07:08.040 --> 07:10.200 The trap is what I call the human trap. 07:10.360 --> 07:13.720 And you may not know what on earth I'm talking about, but we will get to it. 07:13.720 --> 07:18.920 Let's let's now cover the five tricks and the one trap of a genetic AI.