Building TableauTim GPT v1 : Part 1

0:00So many questions, so many things to go

0:03through. That's right, Tableau Team GPT.

0:06Let's get started.

0:07Okay, so to get to this point has taken me

0:10around about a week and throughout that

0:12journey I've gone

0:13through peaks and troughs and I'm here to

0:15essentially cut out all of those and give

0:17you

0:17the shortcut to how this is working and

0:20where Tableau Team GPT will actually go in

0:22the future.

0:23Okay, so the first step is what am I

0:25looking at? Well, I'm looking at a GPT

0:28which is what OpenAI

0:29are calling them that is deployed on the

0:32chat GPT store. This is something that Open

0:34AI announced

0:35last week and as part of that announcement

0:38they actually announced a whole slew of

0:40changes.

0:41One of them was an enhancement to chat GPT

0:42itself giving it the ability to search the

0:44internet.

0:45The other was an increase in the number of

0:47tokens so you can essentially feed it more

0:50text or feed

0:50it more information per query which is a

0:53pretty big change and then on top of that

0:55they've also

0:56opened up this possibility for an app store

0:59like ecosystem and they've allowed people

1:02to create

1:02their own model and this is only available

1:04through chat GPT plus essentially and I

1:07actually had to

1:07shut down subscriptions and signups because

1:11it just got out of hand and basically they

1:13didn't

1:13have enough servers to meet demand so they

1:15've shut it down for a while so if you're

1:17not already in

1:18this you just have to wait a while but I'm

1:20going to show you through what it takes to

1:22kind of make

1:23a GPT and I have to be up front and tell

1:25you that I don't think this was as amazing

1:28as it initially

1:29sounded and I'll get into that in a second.

1:31Now how do you do this? Well it's quite

1:33simple if I

1:34open up my side pane here you can see that

1:37I have some GPTs but if I go to the explore

1:39tab

1:40and push this back so we can get the full

1:41screen you can see that I have some

1:43recently used GPTs.

1:44A friend of mine who runs a podcast called

1:47Damn Kind has a knowledge graph which is

1:49essentially

1:49a GPT that allows you to ask questions

1:52about the topics he's covered in his

1:54podcast which is

1:56sensational and the other ones that you can

1:58see here are ones that are untitled one

2:01that I'm

2:01playing around with at the moment and of

2:03course TableauTim GPT. Let me just show you

2:05what it's

2:06like to create one let's go ahead and hit

2:08create GPT and when we do this we

2:10essentially get a

2:11prompting interface on the left and on the

2:13right we get a preview of our GPT so what

2:15this essentially

2:16is is that if you don't do anything on the

2:19right you basically have chat GPT the

2:21version for the

2:23most recent up-to-date one as of last week

2:25as you enhance the prompt what it's

2:27essentially doing is

2:29configuring the way chat GPT behaves to be

2:31specifically good at answering specific

2:34types

2:34of questions so if I come out of this new

2:37one and I go back to mine and I click edit

2:39you'll see that

2:40the prompt is still blank because we haven

2:42't done anything today but if I go to the

2:44configure tab

2:45it's got the configuration for my GPT so it

2:48starts off with a name pretty easy a

2:50description and then

2:51you've got the instructions and the way the

2:53instruction is built is essentially by

2:55taking

2:55your prompts as you walk and talk through

2:58that chat interface and it summarizes it

3:00down into

3:01a paragraph based on things you've said and

3:03you can also change this yourself so you

3:06can see that

3:06I've pointed it to specific things I've

3:08told it to go and look at certain things I

3:10've given it

3:11certain pages and then as you go down you

3:13have things which are conversation starters

3:15things

3:15that people can ask to give them hints

3:17about what your GPT can do and then came

3:20the bit that I think

3:21excited me massively which was the

3:23knowledge section what you can do with the

3:25knowledge

3:25section is essentially feed it information

3:28and what it does what I thought it

3:30initially did is

3:31it was going to train itself on that

3:32knowledge but it doesn't actually do that

3:34what it does

3:35is it searches the knowledge during the

3:37query so it's a bit like you've gone out

3:39and got a certain

3:40piece of information and you've put it as

3:42the context of your GPT you're going to get

3:44it to

3:45search that information whilst using chat G

3:47PT's own capabilities on top of that it's

3:50going to go

3:50out and search things on the internet and

3:52it's also been updated to April 2023 so it

3:54's going to

3:55be up to date and then at the bottom you

3:57have the capabilities one of the changes

3:59they made was they

4:00made the GPT context understand image

4:02queries and code interpreter in one context

4:06which means

4:06you can be having a conversation with GPT

4:09get into let's say a mathematical problem

4:12use code

4:12interpreter to solve it go back into the

4:14conversation and having had that

4:16conversation

4:16use the information you've garnered and ask

4:19it to create a graphic from that so it can

4:21kind of do

4:21all of that work in one context no need to

4:24kind of switch between tabs and move the

4:25information around

4:26so that sets the context the final thing

4:29that's super important here is this small

4:31bit at the

4:31bottom which is actions and this is really

4:33where the power is because what you can do

4:35you can go and get essentially API systems

4:38that have been set up to work with the open

4:40eye schema so an example they showed off

4:43was getting Zapier or Todoist or a tool

4:46called Retour

4:47to essentially integrate with chat GPT so

4:49you could perform actions whilst in your G

4:52PT so you

4:53can kind of make very specific things a

4:55good example is for example sending

4:57information to

4:57a google doc that isn't a good example of

4:59an action and it's got all the things you'd

5:01expect

5:01including authentication as well so you can

5:04actually get this to authenticate with

5:05systems

5:06that you're familiar with or you can build

5:09your own open ai schema setup and api and

5:11get it to

5:12work you of course need a privacy policy

5:13and everything for that to kind of play

5:15ball because

5:16of course you're going to put this in the

5:17marketplace so you will need a privacy

5:18policy that

5:19people adhere to but that is the

5:21configuration now in the context of tabloid

5:24sim gpt what i was going

5:26to do and what i thought this system was

5:28doing was essentially taking what i was

5:31putting in here and

5:32using it to retrain the model the reason

5:35that's probably not possible because if you

5:38imagine how

5:39many people are doing this they're all

5:40going to be putting their data in it just

5:42costs too much to

5:42retrain the model and everything we're

5:44throwing at it so they might use it for

5:46contextual information

5:47but what does make sense is searching and

5:49sort of refining and honing down the gpt so

5:52it knows

5:52where the context should be and that's kind

5:55of handy so what i initially did was i took

5:58the top

5:5950 videos on my channel between them they

6:01've got about two and a half million views

6:03and i took the

6:05transcript from those and um 10 transcripts

6:08at a time i merged them into one file and

6:11so i've

6:11uploaded these files here but you can see

6:14that it's it's not it's you're going to see

6:16the results

6:16it's not actually that helpful because it's

6:18actually quite a lot of information to

6:20search

6:20bearing in mind some of these transcripts

6:22one of them is like four hours long another

6:23one like two

6:24hours long they're not easy to search and

6:26so i kind of missed the trick here and

6:29where i think

6:29this tool is really good is actually

6:32building smaller more precise gpts so what

6:35i'm going to do

6:36is i'm going to take tablo gpt tablo sim g

6:39pt down a separate path and there is one

6:43there's one huge

6:43detail here which is to do with copyright

6:45and content ownership which i'll come to in

6:48a second

6:49and i'm 100 convinced that you know if i

6:51don't do this to my own content someone

6:53else will so

6:53i'm researching and figuring out how how to

6:56do this the best possible way so that i

6:57come out

6:58with the best version of something that

7:01represents my work but there is a ton of

7:03information out

7:04there that you know i don't own that i

7:05think would be absolutely fantastic maybe

7:07owned by

7:08other content creators maybe owned by a

7:09tablet and i hope they're thinking of

7:11capabilities like

7:12this as well because what i could see this

7:15being really good form is smaller more

7:17capable

7:18capabilities a good example is let's say i

7:20want to learn lods and there could be

7:22something in the

7:22gpt store called um tablo lod gpt and if

7:26you're stuck you can go to it and ask hey

7:29what can i do

7:30with a level of detail calculations where

7:34this gets powerful is if you can take that

7:37and embed

7:38it inside of a dashboard and embed it in

7:39the sort of creator experience you can

7:41actually have that

7:42gpt live alongside the data in the context

7:45and help the user work through these

7:47problems

7:48another example a tablet conference has

7:50finished and you've gone to a few other

7:52sessions you've

7:53taken notes you've gone to the keynote and

7:55you want to create a gpt that helps people

7:57understand

7:58what happened at conference you could do a

8:00tableau conference 2024 gpt and you could

8:03essentially have

8:04a conversation with it you could talk to it

8:05and say hey what happened in the conference

8:07what are

8:07the headline things what are those features

8:09tell me more about them explain how they

8:11work how could

8:12i use it in my context really niche really

8:14sort of small defined concepts and the idea

8:17is you'd

8:17have an army of these gpts doing lots of

8:19different things and i think that actually

8:21makes a lot of

8:22sense but that's not how i set up this

8:24particular system so i've got a huge

8:26problem because with

8:28over 450 videos um already sort of created

8:30here on youtube and the transcripts are

8:33pretty easy to

8:34generate and i have a way of doing that so

8:36i'll get on to that but actually what's

8:38more useful is

8:39the summary of those transcripts so for

8:41each transcript taking the key bullet

8:43points out of

8:43them the key instructions the key how to

8:46use synthesizing that right down so that

8:49the body of

8:50what's actually useful in that video is

8:53actually what's taken and then having done

8:55that um build

8:57a sort of library of information based on

9:00my content and then have this gpt actually

9:02use that

9:03and then further down the road i anticipate

9:05you'll probably be able to train your own

9:07model at some

9:08point so we'll be able to enhance that and

9:09keep going further and further so that's

9:12why i'm going

9:12to take this i'm not going to launch this

9:15um i know people will try and want to do

9:17stuff but

9:18this is this was just an experiment and i

9:20think people got very excited by it i got

9:22excited by it

9:23but in all honesty i don't think it's quite

9:25ready but to show you sort of why i think

9:27that i'm going

9:27to show you sort of what it can do and how

9:29it works and just prove to you that

9:31actually most

9:32of this capability is just coming from chat

9:34gpt4 and that's what's making it great what

9:36it feels

9:37like is they're giving people a way to

9:39brand these gpts as their own when in

9:41reality it's just

9:42another vessel for open ai to get more

9:44reach and sort of get out to more people

9:46what we're going to

9:47do is i'm just going to update this and

9:50make it only to myself and what it should

9:52do is it should

9:53just update everything we're seeing and we

9:55should get a link and i can go and view the

9:56gpt so there

9:57we go and so we're here in tableau tim gpt

10:00let's go ahead and say um how do i use

10:02tableau's latest

10:04feature the reason i asked this question is

10:06because i put some information there about

10:08uh

10:09you know the features including tableau

10:12pulse and so you can see here it's not

10:14actually that good so

10:15it says you actually can you specify which

10:17latest feature in tableau you're referring

10:19to so um i'll

10:20say uh tableau pulse i have given it lots

10:24of transcripts about uh my uh um script so

10:29tableau

10:29pulse a recent edition tableau suite of

10:32tools uh offers a new way to interact with

10:34data get

10:36installed with tableau pulse now i'm fairly

10:37confident didn't search the knowledge that

10:39i have

10:40on this um but this is basically what it's

10:44coming up with so into tableau cloud you

10:47won't be able

10:49to use this on tableau server i'm pretty

10:50certain about that working within slack

10:56utilizing ai for

10:57insights exploring data index i mean i'd

11:00say this is like a seven out of ten in

11:02terms of quality

11:03maybe six out of ten in terms of detail as

11:06well um

11:07and let's see it keeps going it keeps going

11:12and it's not too bad okay yeah um it's fine

11:19it's not

11:20perfect i kind of i kind of see why it's

11:22set up the way it is okay i've opened a new

11:24chat

11:25um how should i prepare for a tableau

11:27interview i know it will do this very well

11:29i actually did

11:29a whole separate video about this just

11:31using gpt4 um what really matters is the

11:34prompt what you ask

11:35and then how you frame the question but

11:38otherwise it's it's pretty damn good um

11:40what are some best

11:41practices for data visualization in tableau

11:44so let's have a go at this so understand

11:48your

11:48audience yeah chart the right chart type

11:50choose the right chart type keep it simple

11:53use color

11:54i mean these are all good i i like all

11:59these but the detail in them is what's

12:04really really tough

12:07right so the interesting thing i haven't

12:10seen today is i haven't seen it search my

12:13knowledge

12:14so either it's got that in context already

12:18or it's um it's just feeding off its own

12:21knowledge

12:22so that's that's another example one done

12:25can you explain tableau's data model um

12:28features

12:28versus joins so this is another example

12:31question that i put in this one's taking a

12:34bit longer it

12:34might just be that it's busy there we go um

12:43they might have changed the interface

12:48because previously

12:48it used to say searching knowledge and now

12:52it's just going right into it so logical

12:55and physical

12:56there relationships logical joins data

13:01blending yeah okay joins physical joins

13:05okay it's got the

13:06context right um because this was updated

13:10in 2023 the open AI gpt4 model was updated

13:14up until april

13:152023 i think it's just scraped the tableau

13:17website this is almost too fast and too

13:19good to have been

13:20from my transcripts definitely um yes

13:22especially this sort of paragraph this is

13:25like a telltale

13:26sign that's come from the tablo website

13:28because it uses some of this terminology

13:31even this exact

13:32framing like um the way things are sort of

13:35laid out kind of feels a bit like it's been

13:38scraped

13:39from the tableau website so that's super

13:42fascinating and so there is a there's like

13:45an ethical question

13:46here because this says tablets and gpt but

13:48how much of this is actually me and how

13:50much of this

13:51is just my brand and gpt4 kind of leading

13:54you to specifically ask questions about

13:58tableau

13:58you might associate its performance with my

14:01content but in actual fact there's no clear

14:04way of knowing how much of this is based on

14:06my content and how much of it is just based

14:08on chat

14:09gpt scraping the entire internet without

14:11people's permission for the record and

14:13using it to their

14:15to their advantage um so yeah i think the

14:17legal the legal the legal sphere and this

14:20is going to be

14:21absolutely wild but thankfully um at least

14:24with my own content this is going to be

14:27super easy to

14:27do and i can play around with it and that

14:29does bring me to the point about content i

14:31think

14:32there are going to be people building gpts

14:36um just by going off and scraping blogs

14:40scraping

14:40uh websites and not really asking

14:42permission so um i think that's just

14:44something to be wary of if you

14:47if you've been making content about a

14:49particular thing just be prepared for

14:51people to just wholesale

14:52copy it the reason i say that is because

14:56this is exactly what tik tok and youtube

14:59shorts mostly are

15:00at the moment you go and look at some of

15:01the most viral content it's not actually

15:03been created by

15:04the people posting the content they've been

15:06clipped from youtube or they've been

15:07clipped

15:08from sections of other videos where those

15:09creators aren't actually getting the the

15:11relevant

15:12recognition so this is a wild wild west and

15:15so it's one of the reasons i'm sort of a

15:17bit reluctant

15:18to just put something out there because in

15:20all honesty this might say tableau term gpt

15:23but i

15:23don't know where it's getting the

15:25information from and you might incorrectly

15:27associate this

15:28content with my content and yeah i'd like

15:30to think there's only one place to go watch

15:33my content

15:34that's youtube or my website and so yeah

15:36for now i'm going to keep working on this

15:39maybe we'll make

15:39some fun gpts based on transcripts based on

15:41things like my crash course maybe we'll do

15:44some more

15:45specific ones and i'll brand them

15:47specifically that way and but i think until

15:49i have a way of

15:50dialing in how much of my own content goes

15:52into this knowledge i'm probably going to

15:54stay off for

15:55a while

15:58you

16:07[ Silence ]