WEBVTT 00:00.200 --> 00:05.440 And just before we take this thing for a drive, let me just mention that I made a little mistake here 00:05.440 --> 00:09.400 when I was changing the tools I left off those brackets. 00:09.600 --> 00:11.480 Uh, and that caused it to crash. 00:11.480 --> 00:15.480 So you may have spotted that, in which case you should have shouted, uh. 00:15.800 --> 00:22.480 I, uh, have now fixed that after a few minutes of messing around and, uh, remember, when things 00:22.600 --> 00:27.120 stop working, the first thing he should be suspicious of is what did you change? 00:27.800 --> 00:30.040 Uh, so now let's run this. 00:30.040 --> 00:36.160 So we run it by bringing up our terminal control and the backtick, and then we move into the fourth 00:36.160 --> 00:39.360 directory, and then we don't do Python. 00:39.640 --> 00:40.320 Uh, we do. 00:40.360 --> 00:47.160 You've run and then app dot Pi to bring up our application. 00:47.160 --> 00:48.000 And here it is. 00:48.000 --> 00:51.680 Here is the sidekick personal co-worker at your disposal. 00:51.680 --> 00:57.560 And I'm rather hoping that what you see is a bit different to this, because I'm loving this project 00:57.560 --> 01:02.760 and I'm fully planning to keep working on it, uh, and to keep improving it in time. 01:02.760 --> 01:08.000 And so don't worry if what you see doesn't match this video, because you should revel in the fact that 01:08.000 --> 01:12.640 it'll have more functionality, more ability, and and if it doesn't, then please help. 01:12.680 --> 01:14.080 Come and help me contribute. 01:14.400 --> 01:18.400 Make some tools, make some features and push them so this can look better and better. 01:18.640 --> 01:21.400 Uh, anyways, now let's set some challenges. 01:21.680 --> 01:26.400 Okay, let's start with what is pi times three? 01:27.640 --> 01:29.000 Let's let it have a think of that. 01:34.360 --> 01:36.000 And the funny thing is that, uh. 01:36.040 --> 01:36.880 Well there we go. 01:37.280 --> 01:40.480 Uh, this is, um. 01:44.560 --> 01:46.040 Well, that's interesting response. 01:46.040 --> 01:49.320 Addressing the previous mistakes about rounding. 01:49.600 --> 01:54.600 So, uh, what we're seeing there, I haven't seen this before, and this is something that we'll have 01:54.600 --> 01:59.080 to, uh, have to, uh, look into, but the. 01:59.120 --> 02:02.320 So we're seeing a value of pi, um, which is great. 02:02.360 --> 02:07.370 And I like the way it's put pi with the Greek symbol and of pi times three. 02:07.530 --> 02:12.610 And we're seeing in the feedback, the final response accurately states that it's that this is correct 02:12.610 --> 02:15.770 and precise, addressing the previous mistakes around rounding. 02:15.850 --> 02:17.090 So I wonder what that even means. 02:17.090 --> 02:23.010 It sounds like, unbeknownst to us, the agent and the evaluator had a conversation about this. 02:23.010 --> 02:24.250 Let's go and take a look. 02:24.290 --> 02:26.610 Wowser there's a lot of information here. 02:26.970 --> 02:28.090 It had it. 02:28.290 --> 02:31.050 It really went to town on calculating pi times three. 02:31.330 --> 02:39.130 Uh, and, uh, see that, uh, it's still we still ended up spending, uh, about 0.1 of a cent, so 02:39.170 --> 02:40.530 a thousandth of a dollar. 02:40.530 --> 02:42.290 It didn't cost us very much. 02:42.410 --> 02:46.570 But, uh, it's interesting that it had it went, went on quite, quite a rigmarole here. 02:46.570 --> 02:47.810 So let's see what happens. 02:48.090 --> 02:53.250 Uh, so this is, uh, so first of all, it went to, um, wow. 02:53.290 --> 02:57.250 It did some looking on online first of all. 02:57.250 --> 03:05.490 And then I think and then it gets to, uh, using, uh, the Python, the Python wrapper and it gets 03:05.490 --> 03:06.210 that. 03:06.210 --> 03:10.050 So it, um, it uses, it uses print results. 03:10.050 --> 03:15.090 So it does follow my instructions because previously it would have got just an empty string there. 03:15.090 --> 03:17.050 So it gets something back there. 03:17.330 --> 03:19.850 9.4477. 03:19.890 --> 03:22.770 Is that what what we see when we, uh, look in the. 03:23.290 --> 03:23.970 Yes it is. 03:23.970 --> 03:24.330 Okay. 03:24.370 --> 03:24.690 Okay. 03:24.730 --> 03:27.290 So it did it did faithfully get there in the end. 03:27.610 --> 03:30.450 Uh, so now we keep going down. 03:30.690 --> 03:34.050 Let's see what happens when the evaluator evaluated this. 03:34.370 --> 03:43.130 So, um, based on this, it says the answer provided by the assistant is incorrect. 03:43.130 --> 03:47.170 The result of this should be approximately that, not 9.425. 03:47.450 --> 03:48.050 Let's see. 03:48.130 --> 03:48.290 So. 03:48.330 --> 03:51.730 Oh yes, it replied approximately 9.45. 03:51.970 --> 03:54.890 The assistant rounded the answer too early or incorrectly. 03:54.930 --> 04:00.770 Therefore, while the intention was to provide an approximate value, the precision was lacking. 04:01.050 --> 04:02.610 And so that's great. 04:02.610 --> 04:04.010 That's this is amazing. 04:04.010 --> 04:10.100 So, uh, Were the, um, the the assistant gave an answer that was too imprecise. 04:10.100 --> 04:14.580 It wasn't to enough decimal places, which annoyed the evaluator that rejected it. 04:14.580 --> 04:19.980 And so then, uh, for some reason, the the assistant decides it's going to go back and run the tool 04:19.980 --> 04:20.860 a second time. 04:21.060 --> 04:22.700 It gets a syntax error. 04:23.340 --> 04:24.700 Let's see that syntax error. 04:24.940 --> 04:25.420 Look at that. 04:25.420 --> 04:27.740 It ends with a curly instead of a closed bracket. 04:27.740 --> 04:29.100 So it makes a mistake. 04:29.500 --> 04:30.940 It uses it again. 04:31.140 --> 04:34.100 This time it gets it right and it gets a good output. 04:34.220 --> 04:36.860 And that is then what it what it sends back. 04:37.140 --> 04:45.660 And now we see the evaluator, uh, and uh, the response is what we already saw in the screen. 04:45.940 --> 04:48.380 Uh, that and that now explains exactly why. 04:48.420 --> 04:52.780 Because the evaluator, of course, sees the whole history of this backwards and forwards, which we 04:52.820 --> 04:53.500 don't see. 04:53.780 --> 04:57.860 Um, and, uh, so, um, yeah. 04:57.860 --> 05:02.980 And if you look at what the evaluator got to evaluate, this is the result of that function. 05:02.980 --> 05:06.420 I wrote that utility function with the user assistant, user assistant. 05:06.420 --> 05:07.980 And you can see that it says user. 05:08.020 --> 05:09.300 What is pi times three? 05:09.340 --> 05:12.500 Assistant uses the tools and then gets that approximate thing. 05:12.540 --> 05:14.940 And the evaluator feedback says that's no good. 05:15.180 --> 05:17.660 The assistant uses the tools and gets a syntax error. 05:17.700 --> 05:23.180 Uses the tools again and gets the accurate number and then gets the value right there. 05:23.900 --> 05:26.060 So this is really fascinating. 05:26.060 --> 05:28.340 What what a journey to see this happen. 05:28.340 --> 05:29.860 It all happened in a couple of seconds. 05:29.860 --> 05:36.060 So we weren't aware that this uh this syntax error this this feedback was happening and we got to the 05:36.060 --> 05:36.980 right answer. 05:36.980 --> 05:39.900 So, uh, that's that's a very, very cool to see. 05:39.940 --> 05:44.740 Okay, I press the reset button, which creates a new instance of sidekick and it rebuilds the graph. 05:44.780 --> 05:46.860 So we've got a completely fresh one here. 05:46.860 --> 05:50.940 And I'm going to paste in a question I'd like to go to dinner tonight in a French restaurant in New 05:50.940 --> 05:51.140 York. 05:51.180 --> 05:58.580 Please find a great French fresh restaurant great French restaurant and write a report in markdown to 05:58.620 --> 06:02.300 dinner, including the name, address, menu and reviews. 06:02.340 --> 06:05.140 Send me a push notification with the restaurant name and phone. 06:05.140 --> 06:06.540 Let's make that tomorrow. 06:06.940 --> 06:07.830 Let's be fair. 06:07.870 --> 06:09.390 It's getting late. 06:09.430 --> 06:12.110 All right, let's, uh, let's give this a try. 06:12.350 --> 06:14.550 Now, there is a risk here. 06:14.550 --> 06:16.190 We will see what happens. 06:16.550 --> 06:18.990 The the risk is that. 06:19.310 --> 06:21.910 And what's cool that it's just popped up a browser window. 06:21.910 --> 06:25.390 It could have used the Serpa API to be searching that way. 06:25.390 --> 06:28.230 Or it could bring up a browser and drive it this way. 06:28.230 --> 06:29.430 So I'm not sure what it's going to do. 06:29.470 --> 06:32.910 And it might bring up the browser, but decide not to use it as it appears to be doing now. 06:33.590 --> 06:34.990 So it's got. 06:35.030 --> 06:35.910 Oh, and it's finished. 06:35.910 --> 06:36.670 It's finished. 06:37.470 --> 06:42.070 Uh, I've uh uh, hang on and see if I've got that pushed. 06:42.070 --> 06:44.310 I have my phone on silent so we won't have heard it. 06:44.310 --> 06:48.830 And I have indeed got a push and it's pushed notified me about Le Bernardin. 06:49.030 --> 06:49.470 So. 06:49.550 --> 06:51.870 Uh oh, it's chosen some snazzy French restaurants. 06:51.870 --> 06:52.830 I'll have, you know. 06:53.030 --> 06:55.790 Um, so, um, let me see. 06:55.790 --> 06:56.990 It's compiled a report. 06:57.030 --> 06:57.670 It's named it. 06:58.470 --> 07:02.990 The report includes detailed information about Le Bernardin, Balthazar and Daniel, such as their addresses, 07:02.990 --> 07:04.030 menus and reviews. 07:04.270 --> 07:07.750 And those, by the way, are phenomenal French restaurants. 07:07.750 --> 07:12.150 But, uh, Le Lebanon is like a three Michelin star or two. 07:12.670 --> 07:14.270 Daniel is two, I think. 07:14.350 --> 07:18.750 And, uh, Balthazar is obviously super famous and popular and celebrity haunt. 07:18.750 --> 07:21.910 So these are all like, uh, top French restaurants for sure. 07:22.110 --> 07:25.510 Uh, it probably found like some website with, with top French restaurants. 07:25.670 --> 07:30.310 Um, it did indeed send me the push notification with the name and the phone number, but we will need 07:30.310 --> 07:35.110 to verify that the phone number is accurate, which we will do right now. 07:35.470 --> 07:41.590 Um, and, uh, we will also go and, uh, uh, check the file. 07:41.710 --> 07:43.390 Let's do that right away. 07:43.750 --> 07:44.070 Okay. 07:44.070 --> 07:45.750 Well, let's start with the phone number. 07:45.750 --> 07:48.310 So I've got the my push notification in front of me. 07:48.310 --> 07:50.110 The phone number for Lebanon. 07:50.150 --> 07:51.910 According to this push notification is 07:51.910 --> 07:58.550 (212) 554-1515. 07:58.590 --> 07:59.510 Let's see what comes up. 07:59.510 --> 08:01.550 If we Google that, we get Lebanon. 08:03.070 --> 08:04.190 Fantastic. 08:04.190 --> 08:06.990 So well done our sidekick. 08:07.230 --> 08:10.750 And now we need to see whether or not we've got a file. 08:10.750 --> 08:17.550 So if we go back to cursor and look at the file in the sandbox, there is indeed a file called dinner. 08:18.470 --> 08:23.190 Let's bring that up in a preview mode and let's get rid of the terminal. 08:24.750 --> 08:25.710 Here it is. 08:25.710 --> 08:28.070 Dinner report French restaurants in New York City. 08:28.110 --> 08:33.190 Le Bernardin highlights and Daniel and a summary. 08:33.190 --> 08:34.070 There we go. 08:34.430 --> 08:40.830 So that does seem to be a successful report, nicely formatted, nicely written to my file system. 08:41.150 --> 08:44.710 And I know from from trying with this that I can then go back. 08:44.750 --> 08:46.510 I can then ask it to flesh out one. 08:46.550 --> 08:48.270 Well let's do let's do a little bit more. 08:48.310 --> 08:52.750 Let's go one step further and then I'll leave with you to experiment with yourself. 08:52.790 --> 08:53.110 Okay. 08:53.110 --> 08:54.430 So I'm going to add in here. 08:54.470 --> 09:04.270 Please update the file to only contain information about Le. 09:06.670 --> 09:07.750 Bernardin. 09:08.830 --> 09:09.430 Oops. 09:09.630 --> 09:10.990 No, I didn't get that right. 09:12.920 --> 09:19.000 and include as much information as possible. 09:21.560 --> 09:22.240 Retype it. 09:22.240 --> 09:22.800 BR. 09:22.960 --> 09:23.400 BR. 09:24.200 --> 09:24.760 There we go. 09:25.240 --> 09:27.160 Get as much information as possible. 09:27.600 --> 09:27.960 Um. 09:30.280 --> 09:41.360 And include maybe that's going a bit too far and include um, um, more details including, um, some 09:42.320 --> 09:52.240 extracts of reviews, a summary of reviews, a summary of reviews, and some menu items. 09:54.960 --> 09:56.560 Okay, let's give this a shot. 09:56.920 --> 09:57.400 Go. 10:02.440 --> 10:03.880 Let's give it a few seconds. 10:07.880 --> 10:08.560 Okay. 10:08.600 --> 10:10.000 It says that it's been doing that. 10:10.040 --> 10:11.760 It also says it's sent me a push notification. 10:11.800 --> 10:12.680 See if that's true. 10:13.280 --> 10:13.760 Yes. 10:13.760 --> 10:14.400 It has. 10:14.800 --> 10:15.800 Uh, okay. 10:15.840 --> 10:23.080 Let's go and have a look in here, and we'll just re load this open preview. 10:24.240 --> 10:24.600 Hang on. 10:24.600 --> 10:29.120 Maybe I need to, um, close the screen and open it up again. 10:33.800 --> 10:34.920 Try this one more time. 10:34.920 --> 10:35.600 Hold on. 10:37.960 --> 10:41.320 It is indeed now just on Le Bernardin. 10:41.400 --> 10:41.840 It has. 10:42.080 --> 10:47.080 So it updated a file on the on the drive in its sandbox, which is pretty cool. 10:47.240 --> 10:49.040 It updated it with more details. 10:49.040 --> 10:53.080 The cuisine menu highlights ambience and summary of reviews. 10:53.120 --> 10:54.200 Impeccable service. 10:54.240 --> 10:55.280 It is three Michelin stars. 10:55.320 --> 10:56.120 So is, uh, right. 10:56.160 --> 10:56.480 Good. 10:56.680 --> 10:57.040 Uh. 10:57.240 --> 11:00.320 And uh, yeah, it's got some menu highlights too. 11:00.480 --> 11:02.720 So it has improved the report. 11:02.720 --> 11:07.480 But the thing that I really want to show you is that it knew how to update a file that it had written 11:07.480 --> 11:12.120 before, and I also it shows that it understood the memory because I just said the file, I didn't name 11:12.120 --> 11:12.240 it. 11:12.240 --> 11:18.850 And of course it remembered that it had been called dinner, which it was getting from Landgraf's checkpointing. 11:19.010 --> 11:20.490 So there you have it. 11:20.490 --> 11:25.130 There is our sidekick in action, uh, doing plenty of things. 11:25.130 --> 11:27.730 And, uh, and of course, it can do a lot more. 11:27.770 --> 11:29.370 You can keep experimenting. 11:29.370 --> 11:31.770 It can do a lot more just with the tools it's already got. 11:31.810 --> 11:34.370 It can build files, it can write reports. 11:34.370 --> 11:38.930 The specific thing I did for a work thing is I had it go and search and browse a number of different 11:38.930 --> 11:41.530 sites, compile them and produce a report. 11:41.530 --> 11:42.650 And it did all of that. 11:42.890 --> 11:44.530 Uh, and it was really impressive. 11:44.530 --> 11:50.570 If you want it to make PDF reports, you can add a tool that just converts markdown to PDF. 11:50.610 --> 11:51.410 That's the trick. 11:51.570 --> 11:58.010 Uh, because the these, um, these, uh, llms are great at generating markdown and they would have 11:58.050 --> 12:01.290 a harder time building a PDF, although they could do it, but it would be harder. 12:01.290 --> 12:07.330 But make markdown and you can write a tool that just uses a Python library to convert markdown to PDF. 12:07.490 --> 12:10.570 And so then you can build those kinds of reports as well. 12:10.810 --> 12:12.450 And it can do real work. 12:12.490 --> 12:13.450 It really can. 12:13.450 --> 12:15.010 And you should put it to the test.