WEBVTT 00:00.160 --> 00:03.200 And this may be more more theory than you thought you were signing up for. 00:03.200 --> 00:04.840 But this is going to be really useful. 00:04.840 --> 00:09.720 This is going to give you some valuable intuition when we actually get to building agent systems. 00:09.760 --> 00:10.040 Okay. 00:10.080 --> 00:12.520 The third one is a simple one chaining LMS. 00:12.720 --> 00:18.760 Look, you you probably know this one already, but you can write a complicated prompt to an LM, like 00:18.760 --> 00:20.680 come up with a puzzle and then solve it. 00:20.680 --> 00:25.080 And sometimes it's a good thing to do because it gives the LM a lot of flexibility. 00:25.120 --> 00:26.880 We like to use the word autonomy. 00:26.920 --> 00:31.680 It gives the LM the ability to choose to go in different directions, to come up with two puzzles and 00:31.680 --> 00:33.160 solve them both if it wants to. 00:33.360 --> 00:39.160 You know, you're giving it quite a broad remit, but sometimes you want it to always have one puzzle, 00:39.160 --> 00:40.200 one solution. 00:40.200 --> 00:44.680 You want to be very careful to make sure it's a hard puzzle, and you want to frame that well, in which 00:44.680 --> 00:48.480 case you could divide this into two separate LM calls. 00:48.480 --> 00:53.360 The first LM call is just come up with a puzzle, and you'd maybe refine that prompt. 00:53.400 --> 00:57.160 You'd add in some some details, some specifics. 00:57.160 --> 00:59.600 So it gets really good at coming up with a puzzle. 00:59.760 --> 01:05.000 And then when it responds with the puzzle, you can you can build a second prompt to an LM and just 01:05.000 --> 01:07.400 include that puzzle in your second prompt. 01:07.400 --> 01:09.560 You can say, hey, I'd like you to solve a puzzle. 01:09.560 --> 01:10.400 Here is the puzzle. 01:10.400 --> 01:11.880 And then you put the puzzle in there. 01:12.160 --> 01:15.640 Now when you do that, we would draw it graphically like this. 01:15.640 --> 01:20.200 We would show it as if it's like one LM call, calling another LM call. 01:20.200 --> 01:24.960 But of course, what's really going on is we're just making an alarm call, taking the output and using 01:24.960 --> 01:26.480 that in our next prompt. 01:26.480 --> 01:30.520 And so and so that that's how we would show it is these two different workflow steps. 01:30.760 --> 01:32.920 And that's what we would call chaining LMS. 01:32.960 --> 01:39.160 It's kind of obvious, but you use it a lot to put more control over your separate calls and to test 01:39.160 --> 01:40.240 them individually. 01:40.280 --> 01:40.800 Okay. 01:40.840 --> 01:44.560 And the fourth one everyone's favorite is tools. 01:44.600 --> 01:47.760 Tools is something that that feels super magical. 01:47.760 --> 01:50.040 And it is in fact really mundane. 01:50.200 --> 01:55.560 Uh, and again, this may be one of those things that, you know, but there's no harm in emphasizing 01:55.560 --> 01:56.000 this. 01:56.040 --> 02:02.720 Let's suppose we have an LM set up where the user asks a question like, what's the stock price of Google? 02:03.000 --> 02:06.320 And we want to call an LLM with that. 02:06.320 --> 02:12.160 So we've got some code or some software or a product like like an A10 and it's sitting there. 02:12.200 --> 02:18.880 It takes that user question and it's going to make a call to an AI, to a large language model to get 02:18.880 --> 02:19.240 that. 02:19.280 --> 02:22.040 It's going to, to, to make a call to an LLM. 02:22.320 --> 02:26.480 And it's an LLM that has been equipped with tools, as we would say. 02:26.480 --> 02:28.560 It's got it's got these extra powers. 02:28.560 --> 02:35.880 It's been told you have the ability to look up stock prices using this product, this product that we 02:35.880 --> 02:37.880 saw last time called Market Stack. 02:38.160 --> 02:40.920 And it's able to make an API request. 02:40.960 --> 02:47.160 It's able to connect to the internet, make an API request and find out Google's stock price. 02:47.160 --> 02:53.080 And then in its response back to us, it includes it says, hey, the price of Google stock is whatever. 02:53.280 --> 02:55.640 And, uh, it seems magical. 02:55.640 --> 03:01.200 And we know that Llms are responsible for taking an input and generating output tokens based on pattern 03:01.200 --> 03:01.960 matching. 03:01.960 --> 03:07.960 So at what point in all of that did this LLM suddenly have the ability to to say, oh, I'm not going 03:08.000 --> 03:09.080 to generate tokens. 03:09.080 --> 03:12.000 Instead of that, I'm going to connect to an API. 03:12.120 --> 03:13.600 I'm going to connect to the internet. 03:13.600 --> 03:14.760 Go and do something. 03:14.800 --> 03:15.720 Get back the results. 03:15.720 --> 03:17.880 And that's what I'm going to put in my output tokens. 03:17.960 --> 03:18.880 How did that happen? 03:18.880 --> 03:21.440 How did we equip an LLM with these tools? 03:21.440 --> 03:23.920 And of course you already know the answer to this. 03:23.920 --> 03:26.280 I expect it is more mundane. 03:26.280 --> 03:31.480 It's not that GPT has this power to stop generating tokens and connect to the internet instead. 03:31.600 --> 03:36.240 No, no, what really happens is this very much more simple the. 03:36.280 --> 03:43.880 Net code when it prompts the LLM, when it's building this message that's going to GPT that it's going 03:43.880 --> 03:44.960 to try and complete. 03:45.160 --> 03:48.560 It doesn't just say what's the stock price of Google. 03:48.560 --> 03:51.800 What it says is, look, you can do two things. 03:51.800 --> 03:58.120 You can either respond with an answer, tell me the answer to my question, or if you wish, you can 03:58.120 --> 04:05.800 respond and say, I need you to run the following tool on my behalf, and then I will or will call you 04:05.800 --> 04:08.720 a second time with the result of doing that. 04:09.200 --> 04:12.050 The tool you can call is looking up stock prices. 04:12.210 --> 04:16.450 The question from the user is what's the stock price of Google? 04:16.690 --> 04:20.450 And the LM will get that whole input sequence. 04:20.450 --> 04:23.090 We've carefully packaged this input sequence. 04:23.090 --> 04:30.250 And the output it will give will say I want to to find the stock price of Google, please use the tool. 04:30.370 --> 04:35.810 And then you call it a second time with all of that and with the tool and it responds with the price. 04:35.810 --> 04:38.330 So it's just a clever trick. 04:38.370 --> 04:46.010 It's a conjuring trick by by calling an LM and by massaging the input to explain the tools that it can 04:46.010 --> 04:51.330 use, and getting that output and interpreting it, and then doing stuff and calling it a second time, 04:51.330 --> 04:57.970 we're able to give this kind of illusion that an LM is able to run tools, but it all comes down to 04:58.010 --> 04:59.090 clever prompting. 04:59.090 --> 05:02.610 Like so much with Llms, it all comes down to clever prompting. 05:02.810 --> 05:05.170 The use of tools is a is a trick. 05:05.170 --> 05:11.850 And just in case you don't believe me, I always love showing this screenshot here, which I did in 05:12.090 --> 05:15.690 ChatGPT and you can do it to read this prompt. 05:15.690 --> 05:17.490 I just wrote this in ChatGPT. 05:17.850 --> 05:20.170 You are a support agent for an airline. 05:20.170 --> 05:23.610 You answer user questions, just respond. 05:23.650 --> 05:30.050 Only use tool to fetch ticket price for London to retrieve the ticket price for London or for any city 05:30.050 --> 05:31.410 that you choose to name. 05:31.570 --> 05:33.250 Here's the user's question. 05:33.250 --> 05:34.370 I'd like to go to Paris. 05:34.370 --> 05:38.130 How much is a flight and look at what ChatGPT responded to me. 05:38.170 --> 05:41.290 It responded use tool to fetch ticket price for Paris. 05:41.490 --> 05:48.850 It just responds with that and hopefully that crystallizes for you this this trick that is tool calling 05:48.850 --> 05:50.370 right here in ChatGPT. 05:50.410 --> 05:54.090 And you can go and do it to yourself to get that first hand experience. 05:54.090 --> 05:57.690 And then the fifth trick is this idea of an agent loop. 05:57.850 --> 06:06.050 This is this concept that you could have an LLM that is repeatedly called and is able to run tools continuously 06:06.050 --> 06:09.930 until it's finished, and it's achieved its goal and it's done. 06:10.090 --> 06:16.010 And in doing so you get this, this behavior where it appears to go off and do some work to achieve 06:16.010 --> 06:16.570 a goal. 06:16.930 --> 06:18.490 So here's an example. 06:18.490 --> 06:20.930 Supposing we had this prompt to an LLM. 06:21.050 --> 06:24.090 Your task is to find the current value of my portfolio. 06:24.330 --> 06:27.730 You are able to number one retrieve my portfolio. 06:27.890 --> 06:31.850 And number two you can look up share price just by responding. 06:31.890 --> 06:33.850 Use my tool to retrieve portfolio. 06:34.050 --> 06:36.530 Use my tool to look up a share price. 06:36.810 --> 06:39.290 So this is given to it. 06:39.290 --> 06:41.050 And how does it complete? 06:41.050 --> 06:45.250 What are the tokens that it will put as the most likely next tokens? 06:45.290 --> 06:50.130 Well, if it's given this kind of prompt, the most likely thing it would want to do, given that its 06:50.130 --> 06:55.730 task is to find the current value, it's going to say, okay, I'm going to use tool to retrieve portfolio. 06:56.010 --> 06:56.370 Fine. 06:56.410 --> 06:57.450 Makes total sense. 06:57.690 --> 07:03.450 And now what we do is we take that completion and we put it into the prompt. 07:03.490 --> 07:05.330 We put the prompt, the same prompt as before. 07:05.450 --> 07:08.010 We add in use tool to retrieve portfolio. 07:08.010 --> 07:11.290 And then we add in the results of actually doing that. 07:11.330 --> 07:12.330 We go off and do it. 07:12.330 --> 07:13.610 We retrieve the portfolio. 07:13.610 --> 07:17.050 And let's say that the answer is three shares of Google stock. 07:17.170 --> 07:19.130 So we just put that in there in the prompt. 07:19.130 --> 07:20.770 And this is the second prompt. 07:20.770 --> 07:22.810 We send the LM in a loop. 07:22.810 --> 07:23.690 We just keep going. 07:23.730 --> 07:26.570 We send this and it then says okay. 07:26.610 --> 07:31.610 The most likely way to complete this input is use tool to look up share price Google. 07:31.890 --> 07:34.050 And we say fine, we take that. 07:34.050 --> 07:39.450 We go and look up the Google share price and we come up with a new prompt which has use tool to retrieve 07:39.490 --> 07:40.130 portfolio. 07:40.170 --> 07:45.610 Three shares of Google use tool to look up share price of Google share price is $100 or whatever. 07:46.010 --> 07:48.530 And we shove all of that in the prompt. 07:48.530 --> 07:53.690 And this entire thing on that's written here goes in one prompt to the LM. 07:53.930 --> 07:55.410 Every prompt is stateless. 07:55.410 --> 07:57.330 It has the full conversation so far. 07:57.730 --> 07:59.650 And now we continue our loop. 07:59.650 --> 08:00.730 We call it again. 08:00.930 --> 08:02.290 This time it responds. 08:02.290 --> 08:05.210 The value of your portfolio is $300. 08:05.250 --> 08:06.770 That's its final response. 08:06.770 --> 08:08.410 It doesn't call any tools. 08:08.570 --> 08:09.530 It's done. 08:09.530 --> 08:10.850 And there you have it. 08:10.970 --> 08:18.850 That is calling an LM in a loop with tools to achieve a goal that is a genetic AI.