Introduction

I'm going to try and do what you're not supposed to do, but a proper live demo today. So I want to try and give you a little bit of an idea of how GPTs can be used.

So how many of you in the room already have a paid ChatGPT subscription at this point? There we go. If I had asked that question three months ago, that would have halved the hands, which is an interesting change already.

How many of you have built a GPT? OK, getting somewhere. OK, so interesting. We'll have a few here.

Understanding GPTs and Their Applications

A GPT is basically a pre-personalized version of ChatGPT. It's the same underlying model.

For those of you that have the paid version, you will be familiar with custom instructions, where basically you could give it this context that it would remember every time that you talked it through.

Previously, before GPTs, there was this whole custom where people would share chats with each other that were preloaded. I don't know if you had that.

I definitely had that in company where you would have trained basically a specific chat on, I don't know, an HR advisor or a marketing copywriter. And then you would share those around, and they would have the context of the chat that you had said it before and go through.

Now, the more... The better way to do this is to build a GPT that is specifically tailored to a specific use. And then basically, it's able to answer over and over within that context.

Exploring GPTs

OK, so at the very start here, so when you go into Chad GPT, you can see here you've got explore GPTs. Actually, I can go here more. You can see that I have quite a few of them that I built.

So here you can see the GPTs that are actually part of the GPT store. So when you're building a GPT, you can publish it in a public way so that other people, anyone really in the store, can reuse that GPT. many GPTs today that you can look at from other people, tutor me, murder mystery that are featured at the moment.

You can see the ones that are created by your company. So this only works if you have a Teams account. And I'll get into that in a second.

But in order to create one yourself, well, I mean, you can look at your own GPTs if you click on my GPTs. And then if you want to create one, just click here. You have a simple interface.

Building a GPT

Actually, most people will start with this. So there is a GPT builder. For anyone that has never built a GPT before, this is literally...

ChatGPT building a GPT, which itself is built on a GPT, by the way. I don't know how many people know that bit, but the GPT creator is actually a GPT under the hood, which is an interesting little fact. So what that does is basically it asks you questions, and you can literally just answer the questions.

And as you answer the questions, it guides you through what the GPT is that you want to build. So do you want to build a GPT to help you optimize marketing copy? It helps you do that.

Do you want to build a GPT to help you brainstorm? In this case, I'm going to go straight into the Configure tab. Basically, if you go and answer these questions, all it will do is it will try to automatically pre-fill the Configure tab for you.

And the Configure tab is similar to what we had before, which is the custom instructions that you might have used previously. And so you have just a few things. You have the name and the description, which do absolutely nothing.

So it's just for you to be able to distinguish which GPTs you've built. Then you have your instructions. So what do you want this GPT to do for you?

And then you can have some conversation starters, which I haven't found that much use for. But it's literally just if you are publishing something publicly and you want somebody else to understand, wait, what type of questions can you ask this GPT, you can have some conversation starters so it helps other people understand what your GPT does. as you're sharing it.

Optimizing GPT Structures

So I have a structure normally that I use for GPTs. Actually, I'll just go here. So it's fairly simple, but the GPT structure that I use

is this one can you i hope you can read part of it i guess i can i can't enlarge it i am going to do it do that i wonder if i can just there you go just all the heading okay so um there are a few things that i do to build a great GPT. One is I always start with this one sentence at the start.

So this is actually, as Michael was talking about before, the context length. You can build it as big as you want.

But one of the things that does happen is that, depending on the model that you use, but in case of GPT-4, which is what I'll be using based on ChaiGPT, everything that you put at the start of a prompt and everything that you put at the end of a prompt is respected more often than everything you put in the middle of the prompt. And so if you want to make sure that something is actually properly respected and that the that the GPT does what you want it to do, it is more likely to execute what you want if you put it at the start or at the bottom. And it more often forgets what you put in the middle.

The way that I try to combat that is that at the start and at the end, I tell it to look at what's important in the middle. And this actually works, by the way. So you can see here, basically, I say, follow the process and respect what's important.

I give it a persona always. So the idea of a persona is you basically restrict how the GPT is going to answer. Again, anyone that would have done kind of more advanced chat GPT prompts to start with would have had this persona bit. It's like, imagine you are the best salesperson in the world. How would you answer this thing? Imagine you are an expert engineer. How would you answer this thing? That's the concept of a persona. A good way to imagine this is if you had a person, if you had one person in the world that could answer your query or could help you with your query, what type of person would that be? That would be how you fill in the persona.

The goal, simple. What is the actual goal of this GPT? What do you want it to do? any context that might be helpful for the GPT to understand as it is trying to work through your query.

Then you have a process, and the process can be fairly, and I'll give you an example in a second. The process can be a multi-step process. The way that I do this is basically number them one, two, three, four, five, whatever the number of steps are, for the GPT to run through whatever task you want to give it.

And then important, this is all the, when I say usual stuff, but anyone that isn't using this yet, I would recommend you definitely should. So weird things of these models, but absolutely true. Ask it to take a deep breath before answering, giving it a tip when it gives you good answers, emotionally charging it if it gives you bad answers. And then one actually the most important one I'd say is the thinking step-by-step. So this is called the chain of thought reasoning. So the idea being that it first outlines what it how it thinks before it actually gives you the answer. And this is just a particularity of large language models, which is that basically your answer, the start of its answer is part of the query. So it's basically what it's doing. It's always only predicting the next token, which is somewhere in between a few letters and a word. And so if it first outlines its own reasoning, its own reasoning becomes part of the final answer it gives you. Interestingly enough, it hallucinates less, it gives you better answers, it solves a whole bunch of problems, and it actually becomes much more powerful.

So this is a template that I use. In here, I'm literally just going to copy paste this. I'm going to say demo GPT Toronto.

Live Demonstration of GPT Creation

And before I go and create the one that I obviously have as a fallback, is there any particular task that anyone in the audience would like me to try and automate? Any idea of a GPT we can build live? Yes?

Audience Participation

Convert an email into a Shakespearean font. DAVID MALANI- Love the idea. Not exceptionally practical on a day-to-day basis. If there's nothing else, we'll do that one.

What is it? Is there something that people are like, hey, I have this thing every day in my job that I would love to be able to automate? Yeah?

I'd like you to search environmental standards when you want to build a garden house. Because you have to search all the time. DAVID MALAN- It's a very good one. Probably too long to build live in five minutes.

It's actually possible, by the way. Yes, one more. Give me a brief of an email, like three bullet points. OK, interesting. So that's one that I can work with right now.

And this is obviously live demo. I'll give you a few examples of one that I've spent more time actually developing myself. Or meeting notes. Yeah.

Building an Email Summarization GPT

Sorry? Or meeting notes if it's longer. Yeah, well, I'll do an email thread. So I'm going to say, imagine you're the world's best assistant.

um expert at briefly summarizing email threads into only what's important right the goal is be brief but don't forget anything that's important um actually wait produce a very brief summary. That's the goal. Produce a very brief summary is the goal.

Context, you will be given an email thread to summarize. OK, so here we're going to go through the process. I'm going to say, one, what would be the process to summarize email thread? OK, well, read the email thread to

Produce a brief summary, three. Evaluate your, and this is another trick, your own summary from zero equals to extremely brief, oh sorry, no. lengthy and containing non-essential information to 100 equals to extremely brief, but not missing any critical information for I'm going to say here, rewrite your summary so that it rates 100 out of 100.

Anyone that doesn't know this trick yet, it's another really interesting one. If you get the model to rate itself and then produce a better version based on the rating that it produced itself, it gets extremely, it gets much, much better answers going forward. So you will actually see that live in the one that I'll demo in a second as well. So and then I'll use this.

DAVID J. Exactly, yes. OK, so I'm going to use this. Wait, will we continue? DAVID J. Yep.

Is there a reason why you asked it to imagine yourself as the best assistant instead of telling me? DAVID J. No, no. You can say. Actually, that is a better version. The general rule is, indeed, be as brief as you possibly can be while conveying the information.

So that's actually a better version. This is my reflex of how I used to do it with ChatGPT. But yes? Another question?

Sorry, I thought you had it. OK. Good. So I'm going to try this.

Now, huh. Dangerous of me switching through to my email in a live demo. Did not think this one through. OK.

So here, I've got a thread, which is actually with FutureFit and guys that have helped us get this venue. And I'm going to try. This is the thread. Very simple.

And we'll see what it does. OK, so here, and this is just a test framework. So I haven't even saved the GPT yet. But so it's given a summary.

It's rating itself. So it's going through the process. And it's rewriting its own summary in one go. So previously, this would be a flow that I would go through with chat GPT.

I would ask it a question. I would then ask it to rate itself. And I would then ask it to rewrite the thing. You can now do this with the GPT.

Further Applications and Benefits of GPTs

So there you go. The one I wanted to show you, just so you have a real, yep, of course. Can you use that same technique of the scoring and asking it to reevaluate itself with more creative, like write a song, write song lyrics that are this? You can use it with literally any answer in any scenario with GPT.

And so you can do it in two ways. You can ask it to rate itself without giving a framework. And then it will figure out what is the framework to try and rate itself. Or you can ask it to rate itself based on a framework that you give it.

Now, what is interesting is if you ask it to rate itself without a framework but with chain of thought, So you tell it, outline your reasoning step by step before giving me an answer. It will actually tell you the framework that it's going to use to rate itself before it actually rates itself, which is really interesting, because then you can look at the framework it provides to rate itself, and you can say, actually, re-evaluate your rating. Don't use this criteria, but use another criteria instead, because it has actually given you the thing.

So you can then go back and forth on what is the right framework to rate what is a good reply to begin with. Now, I did prepare another email thread here, which I had anonymized. See, when I prepare, I can do this stuff. And I'm going to show you here.

is this one. So this is a GPT. I will actually run you through the instructions. So very simple.

Persona, best sales person in the world, has a process. Read the email exchange, consult the documents that I gave it. So I uploaded a bunch of documents on who we are as a company, what we do, why we do it, and so on.

Then enhance your understanding of the entities involved by doing an internet search. So basically, what it will do is it will take the email chain. It'll start researching the parties involved. It'll then rate my suggested reply in the email thread.

So imagine I have a draft email with a thread that came before it. It'll rate my reply and suggest how I can improve it and then explain why. So I use this on almost every single important sales email that I send out. It rereads the thread, gives me the context.

It points out, these are things that you can improve. If it's wrong, I don't use it, or if I don't agree with it. But I have yet to find a single time where I used it where I didn't change the email that I then sent out, which is an interesting one to start with.

So very simple here again. Because I've encoded this into a GPT, I can just say, this is the thread. OK. So now it's saying, OK, here the reply is bad.

It's giving me suggestions on how to improve this particular email. then i can go back and forth so actually what i end up doing with this gpt i tweak my suggested draft and i say it's like what if i did this one instead it will re-rate it and it will again tell me actually in this case you have these other things you might want to improve so this is my personal email assistant that i've been using basically on a daily basis now you can Use these GPTs for a bunch of different uses. And as you saw, I've got like 15 of them at this point.

Versatility of GPTs in Various Roles

I have a strategic thinker, where I uploaded documents on my company. And the only thing I ask it to do is to ask me questions based on a problem that I have. So similar to a coaching style, where a coach doesn't tell you what to do, but asks you questions that help you think through a problem. I have a GPT that does that.

I'm thinking about expanding to toronto do more of these meetups what are the things i should think about or i've got a presentation storyteller which i didn't use today and i should have done a little bit more but basically it takes a proposed outline of a presentation and then creates a story that draws the whole thing together i've got a linkedin post writer very simple it takes one of the talks that we have at the meetup then distills those talks knows that i was at the meetup writes a post saying this was an interesting talk or whatever obviously a different thing condenses it and then shares it on linkedin i've got a meeting preparer it researches both parties tries to figure out it takes an agenda and then tries to figure out what are the meeting the briefing notes that i might want to have as i walk into the meeting You can have a customer support answer drafting improvement, typical one.

Sharing GPTs Across a Company

I imagine anyone that has customer support at a high level that doesn't use this yet should be there. And now, one of the things I think is most transformative is you now have this GPT Teams plan, and GPT Teams allows you to share these GPTs across the company. And so across the company, we have a whole bunch of GPTs. We have a designer GPT.

We have a marketing, copywriting GPT. We now have a Josh GPT, which is basically when I am not available and people need to ask a question, I've tried to build a GPT that takes my thinking framework and at least provides a way of getting people to think through their own problems. And we have them across the company.

The Future of Productivity with GPTs

And one of the things I think is most interesting is that we are moving from a society that was basic, that was people and product.

You have a great product. That can be a differentiator. You can have a great team. That can be your key differentiator.

And in some cases, those overlap. Actually, in quite a few cases, those overlap.

But there's a third category that's starting to be created, which is the productivity platform that you set in the middle. The more these GPTs get created, when people now leave the company, their GPTs stay in the business. So I will still, as a company, be able to produce great copy through the GPTs that people have created as they were there. That is not a category of productivity that we had before.

So as companies now create, there's a new layer of productivity that's getting created through these GPTs that people leave behind them. This is a very early stage. This stuff doesn't always work.

Is it in your employment contract that the GPTs are your property? No, the company's property. That has been in employment contracts all over, just like the code you write, anything that's produced on company time. Indeed.

But that is a whole new category. And it's an interesting one because it will allow companies to compound their acquired value over time, which didn't exist before. So yeah, there we go.

Conclusion

Hope it was useful.