0:00Hey, it's Tim here. Over the weekend, the
0:02chat GPT open AI, the company that runs
0:04chat GPT added
0:06the code interpreter for chat GPT. This is
0:09a separate model from GPT for that we've
0:12been using
0:13or GPT 3.5. If you're using the free
0:15version, it's a separate model that is able
0:17to do mathematical
0:19data analysis because it's using Python
0:21behind the scenes. In this video, I put it
0:24head to head
0:25with Tableau. As ever, let's get stuck in.
0:27Okay, so here we are, we're in GPT four.
0:30And if you're
0:31wondering how to enable this, you have to
0:33go into your settings. And there's a little
0:35beta section
0:35there. When you click on that, you're able
0:37to click on the toggle to enable it. And
0:39then it
0:39appears up here on the top right hand side,
0:41I've actually been using plugins already,
0:43the plugins
0:44have been around for a while. But code
0:46interpreter is a new feature. And that's
0:47available here. And
0:48you have to choose one of the models. If
0:50you use plugins, for example, you're not
0:52able to use the
0:53code interpreter. So these are the plugins
0:55that I've been playing around with, I found
0:56them
0:57extremely useful for various things that I
0:59do. But I can't use those plugins. If I'm
1:01trying to use code
1:02interpreter, at some point, I'm hoping chat
1:04GPT enables the capability to use plugins,
1:07and the
1:07code interpreter, because then you could do
1:09some really crazy things. Okay, so I'm
1:11going to put my
1:12face up here. Now what I want to do is see
1:15if I can do the basic analysis on this
1:17dashboard, this
1:19is the default dashboard that ships with
1:21tablet. And so what I want to do is try and
1:24see how close
1:25can we get to some of these numbers? Okay.
1:28And is the analysis that it's going to do
1:30actually up to
1:31snuff? Is it going to work? Is it reliable?
1:34There's already a couple of glaring issues
1:36with this whole
1:37process. For example, the only way that I'm
1:39able to going to be able to do this is to
1:41upload my
1:42data to open AI. And for most people, that
1:44's already a big no no, you shouldn't be
1:46doing this
1:47work with work data with any data that you
1:49don't trust. I'm doing it with superstore
1:51because it's
1:51sample data. This is the demo data that
1:53comes with a tablet. So we're happy to use
1:56that. So
1:56what I always like to do is to discover the
1:58features that chat GPT uses by simply
2:01asking,
2:01Hey, can I do this thing? So hi, can you
2:04analyze the data for me? Okay. And this is
2:07actually one
2:08of the big things that's been in tweet
2:10threads. And I'm always like, Listen, this
2:12tool is not
2:13going to take over your job, right? Like,
2:16these tools are driven by intent, and they
2:19don't have
2:20the autonomy to come up with that intent
2:22without being given some sort of
2:23instruction. So as long
2:25as these schools don't sort of just go off
2:26and decide to do things, which by the way,
2:28if they did,
2:29the world would be panicking, because then
2:31they could go off and decide to take over
2:33things that
2:34people people, people, people are, you know
2:36, concerned about, but nonetheless, here we
2:38are.
2:38So this is what it can do. What is the
2:39nature of your data? What are the key
2:41variables? Are there
2:42any specific questions you're trying to
2:44answer? Also, if you have the data ready,
2:46can you upload
2:46it here, and we can get started. So I
2:48absolutely do have the data set, I actually
2:50have it on my
2:51desktop. So we'll go ahead, we'll bring it
2:52in, you'll see me drag it in here, and I'll
2:54just put
2:55it inside of the text box. And I'll say,
2:58Okay, here is the data. Okay, let's just
3:01have that
3:02conversation. And you'll see that it's
3:04working. This is the crazy thing. When it
3:07says it's
3:08working, it just says, hey, look, finish
3:10working. And if you actually expand this
3:12box, it's going
3:12to type out a bunch of things, it's going
3:14to tell me that it understands what's in
3:16the data.
3:16And this is already pretty impressive. I
3:18mean, I saw this, and I was like, wow, this
3:20is, this is
3:21pretty interesting. And so what we can do
3:23is we can actually open this up. And you
3:25can see what
3:26it does, it wrote the Python for this. So
3:29import pandas as PD. So basically gives it
3:32an alias,
3:32and then it loads the data using that alias
3:35, so PD read the Excel, this is the file.
3:38And it's
3:38essentially mounting that file that we've
3:41uploaded. And it's able to then, you know,
3:43load up the
3:44headers. And that's what it's basically
3:47using. And then it's this, this is what's
3:49crazy, it's actually
3:50done some initial analysis, it's actually
3:52gone to see like, okay, what can we get
3:55here, here's the
3:56customer information. And so it's gone, got
3:58customer ID, customer name, customer
4:00segment,
4:01it's gone and got the order information.
4:03And it's got that kind of nicely grouped up
4:05. It's got some
4:06country region specific information sub
4:09category. But ultimately, all it's doing is
4:11just listing the
4:12fields. There's nothing sort of, you know,
4:14crazy about this is just I think this is
4:15just the order
4:16the data comes in. So, so far, so good. We
4:19know it has the data. Now we can start to
4:21ask it the
4:21basic questions. So the basic question I'm
4:23going to ask, and I've checked that this is
4:25the right
4:25version of Tableau, the right file, because
4:28each version of Tableau comes with its own
4:30version of
4:30Superstore, I've checked it, so we should
4:32be getting exact numbers. So I'm just gonna
4:34ask it,
4:35what was the total sales in this entire
4:36data set. So let's go ahead and do that.
4:38We'll look past the first steps. What was
4:42the total sales in this entire data set. So
4:46I'm giving it
4:47context. And I'm giving it sort of a
4:49purpose. So let's go ahead, we've given it
4:51intent.
4:52Let's see what it comes up with. There we
4:54go. Now. It's done it in euros. I don't
4:59know if that's what
5:00we wanted. I had it in pounds. And there
5:02was nothing in the file, suggesting that it
5:04was euros,
5:05but 2,938,889. 2,938,889.06. Oh, God, let's
5:14, is it correct to be doing that? I mean, I
5:17'm going to
5:18assume that it's correct. If I just go into
5:20the sheet, you probably want to know, how
5:22do I know
5:22it's correct? Let's go into the sheet. And
5:25let's just right click format. We'll just
5:28go ahead and
5:29say measure values, we'll just choose a
5:32number, number custom will add two decimal
5:35places. Okay,
5:36actually, that's not going to work because
5:37it's the sales value, sorry that I need to
5:39actually
5:40adjust here. So we'll go currency custom
5:42two decimal places. And it's a little bit
5:45small.
5:45So we need to make it larger point zero six
5:47team. So it was it was actually correct
5:50point zero six.
5:51That's what it's got there. Two decimal
5:52places. Okay, we trust it. It works. It's
5:55done the total.
5:56Pretty interesting. Okay. That's pretty
5:58basic. I'm going to assume it can do the
6:00other things. Let's
6:01go ahead and ask it the profit ratio. Now
6:03the profit ratio is more complex, because
6:05what it
6:06has to do is it has to go and do the
6:07calculation for profit, it's only got the
6:09base data, the the
6:11profit ratio is not in this data. So has to
6:13go and figure out the margin, then it has
6:15to aggregate
6:16everything and come to this percentage. So
6:18let's go ahead and just check those fields.
6:20So you can
6:20see here there is no profit, there's a
6:22profit value, but there's no profit ratio.
6:24So let's go
6:25ahead and say ask it, what was the profit
6:30ratio across the entire data set as a
6:35percentage,
6:37I'm being specific about percentage because
6:39I think it does help it sort of get things
6:42better.
6:43And actually, that's how you'd so that's
6:45how the numbers have actually shown. So let
6:46's go ahead
6:47and see how long this takes. And we'll look
6:50at the working as well.
6:52A 12.69 absolutely nailed it. So let's talk
6:56about 69. Let's let's just go back into
6:59this,
6:59I want to see if this is being rounded up.
7:01And I actually don't know about it. So let
7:03's go ahead.
7:04Let's go into the profit ratio. This is a
7:07percentage 12.69.
7:09That is so good. I don't know why that
7:12makes me happy. But this is crazy, because
7:15previously,
7:16this was not the case. Okay. So this means
7:18that for every one euro, it still thinks it
7:20's euros
7:21in sales approximately this amount, this
7:23amount of profit, which is which is perfect
7:26. So okay,
7:27it's done the basic maths. This discount,
7:30this discount is the average discount
7:33across the data
7:34set. So let's go ahead and ask it that.
7:37What was the average discount across the
7:43entire data set?
7:47So this is a difficult one, because the
7:49context isn't always clear. When you do
7:51averages,
7:52you have to like assume something. And in
7:54this case, I think it's done the right
7:55thing.
7:56It's gone across the entire data. And I
7:58think it's just aggregated it all down as
8:01tableau would have
8:01done 10.31. Again, I think this is going to
8:05be correct. I think we can confidently say
8:08this is
8:08pretty easy for this thing to do. And if we
8:11say this is a percentage 10.31 is the exact
8:14answer
8:14the tableau has. So at least that face
8:17value this top level here, easy stuff for
8:20chat, GPT,
8:22I'm not even going to bother with the
8:23others, because those are just sums, it's
8:25going to get
8:26it right. Okay, now we get a little bit
8:28harder. Okay. Now, I want to try and see if
8:32I can get some
8:32of the information that I'm getting in
8:34these charts, because these are sort of
8:36more complex
8:37questions. Okay. And I don't want to just
8:39do that, I want to go in here and pick a
8:40filter. So we'll
8:41go ahead, select central. So we'll narrow
8:44this down to central. And we'll give it a
8:47time range.
8:48Let's go ahead and say that we want to go
8:50from the first of this day. And what I'll
8:53do is I'll
8:53copy these dates into chat GPT when I do
8:56the query. And on top of that, I want to
8:59know the value
9:01for Germany. Okay, so Germany, Bavaria here
9:06. And I think it's, well, this is
9:09interesting. So this
9:10is this the whole no, these this is a
9:12region. So this is a, I need to go in and
9:15be specific about
9:16what exactly this is. This is a region. Yes
9:20. So Germany, province Bavaria is a province
9:24in this
9:24particular concept. So state slash province
9:27. So what I'm going to need to ask chat GPT,
9:30and I need
9:30to kind of be specific here, because I
9:32think it really does want you to be this
9:34specific is I need
9:35to ask it in Bavaria for the central for
9:38the specific date and time. What was the
9:42total sales?
9:42Okay, so this is going to be a little bit
9:44hard. But let's, let's, let's do one thing,
9:47actually,
9:47let's go in here. And if I, if I just
9:51duplicate this as a cross tab, it will give
9:56me this field,
9:58and I can just go out here. And I can, can
10:00I copy this, I can edit the alias, I cannot
10:03spell. So I
10:03want to grab this and just say, Okay, cool.
10:06So go back to the view. And then I say,
10:08Okay, what
10:10was the total sales in central region?
10:16Paste that Bavaria in Germany, between the
10:23dates. Let's go ahead. This is a little bit
10:27of a tricky
10:28one. So I hope I get this right between the
10:32dates, this one. And let's get this out.
10:38This is boring
10:39to watch, I'm sure. But hopefully, you're
10:42still with me here. And that now I'm
10:45missing anything
10:46total sales for german, Bavaria, Germany, I
10:48've got no other open, they don't want to
10:50click my face.
10:51I don't want anything else. Everything is
10:54good. Okay, we can just go ahead here and
10:57say, Okay,
10:57what are the total sales in central region
10:59Bavaria in Germany between these two dates?
11:01Okay. Here we
11:03go. moment of truth. What's it gonna do?
11:07Given it two dates have given it all the
11:10other contexts.
11:11That is like an interaction. That's why I
11:13think this is a more complex question.
11:15There's a lot
11:16to get right in there. And here we go. Now
11:19I see here, this is interesting total sales
11:23in central
11:24region specifically in Bavaria, Germany,
11:27between those two dates was 24,000 euros.
11:30Now, that
11:31doesn't look right. Because did I did I
11:34give it the right data? Oh, no, it's it's
11:37got it right.
11:3824,651. Geez, this. Sorry, if anyone's
11:43religious, I apologize. I'll try and keep
11:46it clean here. That
11:47is that is nuts. Wow. Okay, that is that
11:52just blew my mind away. I cannot believe it
11:56got that right.
11:58Let's look at the working. I mean, I don't
11:59know Python. So I don't know what I'm going
12:01to be able
12:02to do here. But this is absolutely nuts. So
12:04it's gone and got the data. And it's
12:06written what looks
12:08like fairly simple Python here, it's gone.
12:10And basically, it's like it's written SQL
12:12saying,
12:13hey, this is the information, go get this
12:15and figure it out. And by the way, it's
12:17been doing
12:18this all along. So if you're not familiar
12:20with Python, it's basically been sort of
12:23building this
12:23up. So you could also use this as a really
12:25nice way to sort of go backwards and figure
12:27things out.
12:27What is alarming is that, and I think this
12:31is why companies like Tableau are really,
12:34you know,
12:34really worried about this technology is
12:37that the threat isn't this this, the threat
12:40isn't that it
12:40can do this, this is not the threat. The
12:43threat is that the interface for how we do
12:45data analysis
12:46might be changing right in front of our
12:49eyes. Everything I've done has not been
12:51drag and drop
12:52here in Tableau. Just imagine just to build
12:54out that chart to build everything here
12:57would have
12:57taken out more time than typing the
12:59questions. And I'm pretty good at Tableau,
13:01I can confidently say
13:02that it would take me less time to type it
13:04than it would to build what's in front of
13:06me, even if I was
13:08racing and trying to do everything really
13:10as fast as I can. The other thing is that
13:14this little box
13:15here, anyone can use anyone can type into
13:18that it doesn't just take me a data analyst
13:20, it takes
13:21anyone who has a basic skill set of asking
13:23good questions that is an important skill
13:25can use this search box. So in my mind, the
13:28innovation behind open AI, all these AI
13:30models
13:32is not the models, it's actually the
13:34interface that allows anyone to use them
13:37well, that is actually
13:38really hard thing to do in open AI, I've
13:40done a lot of work to train these models.
13:42So this to me
13:43is the innovation Tableau, everyone, Power
13:45BI, all these products, they're racing
13:47towards AI,
13:48you might think they're racing there
13:49because you know, this is the next big
13:50thing. No, they're
13:51racing there because I think they all see
13:54this little box as the biggest threat to
13:56their product.
13:57If they can't get their product into this
13:59box in the next five years, they might as
14:02well give up
14:03right now because someone out there is
14:05doing exactly that. And it's actually alarm
14:07ingly good
14:08at being specific about certain things. Now
14:11, this is where tools like our data kind of
14:14went down,
14:14but they were doing something which wasn't
14:16as good as some of these models. And the
14:17key thing
14:18to remember here is that open AI have had a
14:20massive investment in technology to get to
14:23the
14:23point where they could build the model that
14:25is what we're using here. They've had lots
14:27of money
14:27pumped into them by Microsoft and lots of
14:29other investors to get to where they are
14:31today, to be
14:32able to then ship this model out to
14:34companies, the actual model, the final
14:36model they build
14:37is very easy and portable for them to share
14:39with companies and deploy. The hard thing
14:41is training
14:42it and that takes time, people resources,
14:44and a lot of hardware, but pretty much only
14:46Nvidia makes.
14:47So this is pretty crazy. I just cannot
14:50believe this. So we've seen it answer
14:52questions. Let's see
14:55what it can do with charts. Let's go ahead.
14:57I want to go to a different dashboard just
14:59because I'm
15:00interested in this. Okay, so let's ask it
15:05for a sales by product category, broken
15:09down by year
15:10and month. Okay, so let's do that. Let's go
15:14back to this and say, hey, we don't like
15:17tables,
15:17but we'll ask for one. Hey, can you build
15:23me a table showing sales by product
15:27category?
15:31Category with years as a subheading and
15:40months as columns along the top. So I've
15:46put some
15:47information here about how I want to see
15:49the data and let's go ahead and see what it
15:52does. So it's
15:53going to have to do the processing first,
15:55but then it's going to have to visualize it
15:56. So because I'm
15:57asking for a table, it could print the
15:59table to be fair as text. It maybe doesn't
16:01have to,
16:02but here we go. Let's have a look category
16:04year. Oh, this is interesting. Okay.
16:08Oh, my word. Oh God, it's doing it.
16:14So then the next question is, is this
16:19number correct? 6751 for furniture 2020 Jan
16:25. Let's go
16:26ahead 6751 6752. Oh, I'll give it that's
16:32just a roundup. That is that is remarkable.
16:37That is,
16:37I can't believe it got that right. I know
16:40it sounds so simple. You're probably
16:43thinking,
16:44Tim, this isn't that sensational, like, no,
16:46but it is because all I did was type it in,
16:51and it nailed it, it nailed it. And I'm
16:53comparing against a finished solution. I'm
16:56literally drawing
16:57the gap between you don't miss yourself and
17:00typing the question that is that is
17:02ultimately what is
17:03groundbreaking here. And so far, it's got
17:05everything right. There's not been a single
17:08mistake in this entire pipeline. This is
17:11pretty crazy. Now, it did what I thought it
17:13would do,
17:14which is it didn't actually do any sort of
17:16visualization. You can see here, it's just
17:18literally just done a pivot table. And it's
17:20it's actually printed the result as I
17:21thought it would
17:22here. And then it's just basically gone and
17:25styled it in HTML, which is even nicer.
17:27Okay. And it says
17:28the values in the euros, we know that's
17:30wrong. I could have corrected that earlier,
17:31but we didn't
17:32bother. Okay, it's passed that test. Nice.
17:35Okay. We'll ask for this sales and profit
17:38by product
17:38names. Okay, this is a this is a this is a
17:40difficult one. So let's just grab this. I
17:43'll try
17:44and give it less information this time. So
17:47can it show us sales and profit by product
17:50names?
17:50Oh, this is a this is a really, this is
17:52really difficult. This is all about the art
17:55of typing
17:56the good prompt. So wow, what I'm what I'm
17:59going to do this, what I'm going to do is I
18:02'm going to
18:02put tabular on another screen. And I'm
18:05going to put chat GPT here. And what we're
18:07going to do is
18:08I'm going to type the prompt in front of
18:10you. So you can see what I'm typing. So I
18:12'll say, can you
18:14build a scatterplot? I want it to show
18:19sales and profit by product names with
18:27category
18:29and subcategory breakdowns in rows. Nested.
18:39And then what is on the top? What is that?
18:45What is
18:46that official terminology? Let's just go. I
18:48think this is customer segment. Yes,
18:51correct. So we want
18:52segment and then segment has column
18:59sections. Okay. So this is a slightly more
19:05ambiguous thing,
19:07because I've asked you for a scatterplot,
19:09but I'm talking about rows and things being
19:12nested,
19:12and then the segment is column sections.
19:15Let's go ahead and hit enter. Let's see.
19:17This is this
19:18is tough because my prompt could have been
19:20bad. I don't think it's that bad. Probably
19:23has typos,
19:24but let's let's wait and see. I'm not going
19:25to do the work. I'm not going to show the
19:27working. I
19:27always look like to look at the working
19:30afterwards. Oh, it stopped. So what is
19:33going on here? I think
19:34what it's doing is it's potentially firing
19:36up another thing. Oh, wow. Okay. Yeah. It
19:40got busy.
19:40Okay, so here we go. So what did it do? It
19:45imported matplotlib by pyplot. So it's
19:48imported
19:49a library and it's imported seaborne. And
19:52then it's commented it out. So it has map
19:55scatterplot,
19:56create facets. Seaborne seems to be able to
20:00do some of this segmentation. So hold on.
20:05Okay. And it's created it and then it's
20:08basically kind of pulled the charts and
20:11that's what we've
20:12got below here. But did it do it correctly?
20:16So we've got corporate consumer home office
20:20.
20:20Yeah. Okay. Now, what it didn't do nicely
20:27is this this thing here. So what it did is
20:31it kind of said
20:32each scatterplot is going to be a subset of
20:35the total. So it didn't put it didn't put
20:38them along
20:38the left. That is something that kind of is
20:40confusing. And I thought this would be
20:42confusing,
20:43but it's not done a bad job of giving me
20:45the alternative. And I can still see the
20:47outliers.
20:47So I can see the outliers clearly here in
20:50many ways, it's maybe made it easier for me
20:52to do the
20:53analysis because in Tableau, these would be
20:55in one line. And just to show you what I'm
20:57comparing
20:58against, I'm comparing it against the
21:00bottom bottom chart here. And that is a
21:02pretty interesting
21:03sort of take on that isn't exactly what I
21:05wanted, because I was starting with
21:08something that was
21:09already, you know, articulated. The trick
21:11here is going to be prompt. And I think if
21:14I did a better
21:14prompt, it maybe would get closer to the
21:16answer. Also, I'm not sure if the thing I'm
21:19asking for
21:20specifically possible, you know, having a
21:22table like structure on one column, on one
21:24sort of axis
21:25and then like a section set up on the other
21:27and it got some of that, but it's not
21:29perfect.
21:30And it goes and explains what it's done.
21:32And again, you've always got the ability to
21:35show
21:35the working. So that that is that is truly
21:37remarkable. I don't know, I don't know what
21:40you're using this for. If you're using this
21:41, let me know what you think. There is a
21:45couple
21:45of issues with it, it still can't do all
21:48the all the things you would expect. So if
21:51I go ahead and
21:51ask it, hey, can you plot a map with each
21:58data point sized by cells at the level of
22:04cities in the
22:06data set. So I've sort of asked it in a
22:08really bad way, there's even a typo in
22:10there. Sorry for the
22:12inconvenience, but currently in this
22:14environment, I'm unable to access internet
22:16services, which
22:16includes mapping services such as Google
22:18Maps or any other geolocation service, it
22:21could be
22:21used therefore I'm unable to create a map
22:23plot for your data. Fine. However, I can
22:25guide you to this
22:26using Python libraries, such as Geopandas
22:28and folium. So it's basically saying, hey,
22:31you can
22:31do this in your own environment. And by the
22:33way, this is the code to go and get a
22:35little bit of
22:36geocoding that it does, you see it takes
22:38the city column, and it uses geolocate
22:40using the latitude
22:41and longitude to basically get the exact
22:43location. And then it's able to then import
22:46folium, which
22:46takes that information. And for each row in
22:49the data set, is actually able to pass that
22:51and then
22:52create the map. So this telling you how it
22:54would work, if it had access to that. But
22:56the other
22:56crazy thing you've got to remember is that
22:59chat GPT is firewalled from the internet,
23:01everything
23:01you get back, it's not able to search the
23:04internet, it can only use its large
23:06language model or model
23:08that's built into it for everything that it
23:10does. And in more recent features, they've
23:13had this
23:13ability to browse the internet with Bing,
23:15but they actually have to disable that
23:16because it was doing
23:17things they didn't like. But in Bing, you
23:19can actually get it to go ahead and still
23:21do that. So
23:22in here in chat GPT, it's completely sort
23:24of isolated still. And that's how it's
23:26working.
23:27Anyway, what do you think? I think the
23:29biggest threat to pretty much everything is
23:32this search
23:33box, this search box, what is incredible is
23:36this tool, if I didn't expose the working
23:39every single
23:39time, what I'd be seeing is just answers.
23:42Okay. And you might be wondering, well,
23:45what is the what
23:46is the confidence level of some of this
23:47work? Well, I walked away with pretty high
23:49confidence,
23:50the things that can do it told me and
23:52things that didn't quite get right were
23:54based on the
23:55prompts. And that's another sort of
23:58interesting angle. But yeah, and if I'm
24:00Tableau, I can
24:01absolutely understand why they're racing to
24:03build something like Tableau GPT. Because
24:05again,
24:05if in five years your tab your product
24:08doesn't fit in this text box. I'm certain
24:12something will
24:12come in at your lunch. Anyway, thanks for
24:15watching. I'll see you in the next one.
24:17Transcribed by https://otter.ai
24:26[ Silence ]
Python | AI & ML | Data visualisation | Tool strategy