Introduction to Live Demos and GPTs

I often have the privilege of trying to run live demos. So I preface this with this stuff is still fairly temperamental. Sometimes it works, sometimes it doesn't.

But what I'm gonna try today is give you some ideas of how GPTs can work.

So before I start, how many of you in the room already have a paid version of ChatGPT?

This is very interesting. So I did the same thing yesterday in Toronto. And six months ago, that would have been less than half the room. Now it is three quarters, if not more, of the room, which is an interesting difference.

How many of you have already built a GPT for yourselves? That's still not where it could be. So hopefully you'll get a lot out of this then.

Understanding ChatGPT and Personalized Chats

Before GPTs came out, there was a whole big practice of people sharing specifically trained ChatGPT chats. So you would have a conversation with ChatGPT, you would tell it to take on a particular persona, you would give it some data, and then you would keep it there because you now have gotten that conversation to a point where it was useful to you, and you kind of wanted to use it again the next time around, or maybe you wanted somebody else to be able to use that. and you ended up in the sidebar here with a whole bunch of different chats. Now you can name them before you could not name them either.

That was a practice that was happening a lot. And so you'd have one chat that you had trained as a coach. You'd have one chat that was trained as a coder. You'd have another chat that was trained on something else.

The Advantages of GPTs

And then OpenAI came out with GPTs, and GPTs are basically a better version of that. So you can give it a set of instructions that it will take into account every single time you ask it a question. And that means that you can have specific kind of pre-programmed general purpose transformers that perform a particular task over and over. And so if you start to then think, what are the tasks you perform most frequently, you can have a fine-tuned version of Chan-TPT that specifically goes through those tasks and does so in the most efficient way possible so you have very little work to do.

So for those of you that don't know, you can, when you're in the left-hand side here, you can see you've got various GPTs. I have quite a few of them that I built for myself that I use on a regular basis. You can click on exploring GPTs, chat GPTs, or I mean OpenAI, they have a GPT store now.

So when you're actually building a GPT, you can decide to publish it for yourself, keep it private. You can publish it to your team if you're on a team plan. or you can publish it to the wider world and basically then other people can install your GPT.

Creating GPTs: A Step-by-Step Guide

You can create one by clicking on the button here, create, and then you have two different interfaces. So you've got the GPT builder and This GPT builder is a GPT that helps you build GPTs, for those of you that don't know. So there's an actual GPT under the hood here, basically.

This is a customized version of ChatGPT that asks you questions, the result being that by the end of having asked you questions, it produces a set of instructions that it pre-fills for you. It's nothing more than that. So if you are starting out, you can literally just use this. answer as many questions as you want, and it will build a GPT for you.

So here it would ask you, what would you like to make? And I wanted to open this up. So try and figure out if we have a fallback, obviously, if we need to.

Live Demonstration: Building a Recipe Creator GPT

But what is a particular task that people in this room have that we can try as an actual GPT? Anyone have any ideas? Yes. help making a recipe.

Okay. Interesting. A recipe based on stuff you have in your fridge. I like that.

We can try that. Let's, let's, let's look at that. Um, so I'm going to call it, uh, recipe creator, uh, create a recipe with the, um, uh, ingredients I have in my fridge.

Okay. So first thing, name and description don't matter at all. So you can name it any way you want. It doesn't matter at all in terms of how the GPT works.

Um, and then you give it a set of instructions. Now I have a template of how I look at instructions and I will take this. And I'll just copy that here. All of this is recorded, by the way.

So you will all get an email at the end of this. And you can look at this. You don't have to. I mean, you can take pictures if you want.

It's all good. But you also get access to it afterwards. There are a few things that go into these instructions, the way that I structure them. So one, you will see that I start and I finish the GPT instructions the exact same way.

Does anyone know why this is? No. Yes? Because the AI is really good at looking at paying attention to the beginning and the end, but not very much.

Exactly. Depending on the model that you're using, and GPT-3.5 and 4, actually, they are great at respecting the start of the context and the end of the context, but they actually tend to forget or be more forgetful in the middle of the context. So the way that I solve this is to use the start and the end to tell it to pay attention to the middle. Stupid things like this, like some of this stuff doesn't quite make sense.

This is where it feels a little bit like magic, because it's these things that you just need to know, but they really do work. Now, the other thing is fairly standard stuff that I imagine if you've been using ChatGPD for a while, you should be familiar with. So the idea of taking on a persona is just making sure that you restrict the data set from which it starts to answer you. It improves the quality of the answers.

So a persona could be an experienced salesperson. It could be an experienced coder. It could be imagine. If you're asking a question, so here we're talking about a recipe, it would be a chef, an experienced chef, that I would ask this question from.

If I could ask this question from a human, it would have been an experienced chef. And so I'm going to ask my GPT to take on that persona. I give it the goal, obviously, like what ultimately do I want out of this GPT, any context that might be important. So here I'm going to give it context about I want it to restrict its answers to only what I have in my fridge and nothing more.

And then I have the process. And this is actually, it's an interesting, I mean, it's not something new, but I haven't seen many people kind of use this in this way. You can just give it a series of steps. One, two, three, four.

So however many steps. The more steps you give it, the more likely it is to screw up. So you can have really advanced ones, but I would advise if you can get to your result with fewer steps, then you end up with a more reliable GPT. And they have gotten better at this as well, by the way.

Another thing that happens is every single time that OpenAI releases something, they basically swap out the underlying models for more performing ones, which means that your GPTs end up with much more volatile behavior. So somehow you have a GPT that worked one day and then you look at it the next day and they have a spike in usage and so they swap things out and suddenly it no longer is able to follow any of the instructions that you have, which is absolutely crazy. Or at least if you're relying on them, at this point I am, it's a little bit harder. And then you have the last part here, which is everything that's important.

This is all the stuff that most people will know, I imagine, so take a deep breath before you're answering a query. using chain of thought, so outlining your thinking before you actually answer. So this is just a quirk of large language models. Basically, when they answer, the first words that they use to answer your query become part of the result of the next words that come through.

So the way these models work is they predict the next token, which is somewhere in between a few letters and a word. And so if you ask it to outline its own thinking, it will take into account its own thinking in the answer that you ultimately get, which basically makes the answer much more robust. It reduces hallucinations quite a bit and it gets you to higher quality answers.

And then the other two things, which honestly I have, I mean, I have a hypothesis on, but nobody really knows why, but it has been proven over and over. You tip it, you get better answers. You charge it emotionally saying that something bad is going to happen if it screws up, then you also get better answers. All right, so this is the template that I use.

We're gonna just populate this very quickly.

Defining the GPT's Persona and Goal

So the persona here, you are an expert, experienced chef. I'm gonna say the goal, produce me the most delicious recipe with the ingredient, whoop. i tell you or well with the ingredients i have access to i recipes

Yeah, actually that's an interesting, we could do that indeed, but that would probably go outside of the scope of just the timing of this particular demo. Should not include any ingredients I don't have in my fridge. Simple, right?

So I'm gonna say process one, ask me for the ingredients I have in my fridge. Two, ask me for the type of food I like. Three, produce a recipe with the ingredients I have that I am most likely to enjoy.

Refining the GPT Instructions

Now you can create an image and everything you want. I like the idea with the image now. Now I'm thinking like, shit, actually, that's one that I'm really going to create.

I'm not going to save this because I don't really use this. That one is one that I'm going to create tomorrow. But for the demo here, I'll just keep it here.

Conversation Starters and Ingredient Inputs

So I'm going to ask one more thing. I'm going to say conversation starters. What ingredients do we have to work with?

OK. And the conversation that starts here is just, if you share this GPT with someone else, it gives them a clear idea of, OK, how do I interact with this GPT? Because otherwise, you might just ask it a question that has nothing to do with ingredients. And they're like, well, what do I do here? I'm going to start with this. OK.

Can people shout out some ingredient names? Butter. Butter. What else?

Tomatoes. Tomatoes. Sorry, bell peppers.

Ketchup. Ketchup. Dorians. Sorry?

Dorians. Dorians? Dorians. I have no idea.

I have no idea how to spell that or what even that is. D-U-R-I-A-N. Durians.

Anything else? Spatchcock. Sorry?

Spatchcock. See, this becomes an English test for me now. I'm pretty sure that ChatGPT will actually figure out what the hell this might be because I misspelt it somehow.

Okay, I'm gonna keep it at this. I wonder if it can come up with anything. Okay.

Generating a Recipe with User Inputs

Okay, what type of recipe do we wanna create? Mexican. Mexican, yeah, I was gonna say Mediterranean fusion.

Okay, spatchcock, oh, but it hallucinated the chicken. Oh, wow. Okay, so clearly has, I wonder if it did anything else.

Chilli and ketchup, salt and pepper, no. Okay, it, paprika, cumin, okay, wait, this is, so basically I didn't tell it specifically to not use any spices, but I'm, It's not a bad result, but just so that the example is actually good, what I would do here, recipe should not include any ingredients, spices, or anything else I don't have in my fridge or didn't or specifically tell you about.

Okay, so I made the mistake of not copying the list of ingredients, so we're going to do that one more time. So we had butter, we had tomatoes, we had ketchup, we had bell peppers, we had durian, and then we had spatchcock. I don't know.

And we said Mexican. Mexican Mediterranean fusion. That's literally what we were saying before, yeah.

Okay, so let's see what it does now. Live demo, so. Come on, come on, you can do this.

So the difference here would, yeah, so it is the same thing.

Yeah, exactly. It's saving it, which means that if I wanted to use this every day, and every day I'm just going through, it just gives you an easy reference point.

I'm not entirely sure why it's not. I don't think it's the exact same as saving

I wonder what... Okay, so now it's doing... I didn't even know you could just stop it and just go through in the same chat, but you learn every day. Okay, so now figure out if it does this better.

It still goes into the chicken. Spetch khakis. There you go.

Okay, that makes sense. Okay, so tomatoes... So that's the thought process.

Now it actually goes... So this is... Obviously, because I told it to outline its reasoning first. Now it actually comes through here. So it's restricting itself.

It's still... Oh, wait, optional spices. Okay, well, whatever you have that fits. So at least it's doing this in a better way.

It's basically saying, hey, you can't really make anything with this stuff, so please use spices. Okay, so here we have a full generator recipe.

Everyday Applications for GPTs

So this is an example of a GPT that you could be using every day, and as we were talking about before, you could make it so that you upload a picture, it identifies, so the first step would be ask you to upload a picture, second step would be identify the ingredients, and then you would go through the same steps. And so then you'd have a GPT where literally you just go to your fridge, snap a picture, it gives you a recipe. That's actually a pretty cool GPT that you might use on a frequent basis. There are a whole bunch of different ways that you can use this.

I tend to use ChatGPT much more on a professional basis. So I upload it with a whole bunch of documents on what my company is about, what I'm trying to do. And then it first goes and consults everything that we're doing. And then I use it as a sparring partner.

I have one for a strategic thinker. Basically, I tell it to only ask me questions to help me think through whatever problem I have.

I have one that does a presentation storyteller. So if I have a presentation that lends itself to a story and I want to kind of link it together, I basically give it an outline of a presentation and I tell it, give me a story that links all of these points together in a way that I can present to an audience.

I have a LinkedIn post writer. Yes, I do use that.

I have a meeting preparer, and the meeting preparer basically looks up the people that I'm meeting, does a web search on whoever I'm meeting, then takes the goal of the meeting and gives me a briefing note based on whoever I'm meeting and the goal of the meeting itself so that I can walk into a meeting having the briefing note prepared.

You can use this for customer support. So if you have a, not for each one of them because obviously many of them are kind of quick back and forth, but sometimes you have an important one or a difficult one to answer and you can have one that looks up all your documentation automatically and formulates an answer based on documentation that you have available. All of this stuff fairly easy to build.

GPTs in Teams: Transforming Business Operations

Now where this becomes really interesting is that you have this GPT Teams plan. So how many of you are on the GPT Teams plan in your company? Little, okay, few, but still few. Still not that many.

The reason this is becoming interesting is that basically up until this point, Success of a company who's linked was basically a combination of the quality of your product, you can say quality of your position, market, branding, and distribution, and then the team that you have to actually build the company, build the product, and go further.

What is happening here is that you end up with these GPTs that are getting built by people that are permanent. So you can have a copywriting GPT. We have one in the company where anyone that's trying to write copy for the company does this using the copywriting GPT. so that it is on brand and that it uses the terminology that we want to be used.

I have a Josh GPT, which is anyone that has a question for me about how I think about how we're building the company, and if I'm traveling and I'm here and I'm not available, well, go and ask the GPT first. And if the GPT gives you a good answer or an answer that you think is okay, well, then don't go and bother me about it.

Those are things, I mean, these are just a few examples, but those are things that last even if you leave the company. And so when you think about these platforms, companies are now starting to create this layer of extra differentiation, of extra productivity that simply did not exist before. And so we have an additional lever basically that compounds over time. And I think that is a really, really interesting one.

And so just wanted to leave you with one that I use a lot.

Personal Favorite: The Sales Email Thread Assistant GPT

So my favorite GPT for myself at the moment is i have where is it sales email thread assistant very simple so it follows the exact same format process here i ask it is read the email exchange consult your documents if relevant basically do your research on the different parties that are part of the email chain I give it the full thread, and the last part of that thread is my suggested reply based on the thread. I ask it to rate my suggested reply based on the context as part of the thread, and then I ask it to explain why it gives that reasoning and how it can improve that.

So here, I'm going to take, this is an actual example, although I use ChatGPT to mask all the data so that nobody knows exactly who this was with, unless, I mean, someone who might actually see this video, they might recognize the structure of this particular exchange. If I just go fast enough, they might not read that. I just pasted...

This was just to save, yeah, so the way that I would do this, I would go into my emails, I would draft a reply, and then I just copy the entire thread including my draft. before I send it. I just copy-paste it in here, and hear what it does. So it goes through, it reads the email exchange, it then consoles the document so it has a good idea of what we do.

In this case, it doesn't require any additional information on the parties involved because the thread is fairly large. It rates the reply. So here it says the rating is 75. This is one of the things, like actually going through these GPTs has been a humbling experience because the first times I used this, I got ratings of like 15 in some cases, which is like, okay, my sales emails really need improvement.

And then it gives you reasoning. How can this be improved? Well, you can be more clear. You can highlight differentiators more clearly, use bullet points a little bit more, stronger call to action.

So here the invitation for another call is good, but it could be stronger by suggesting specific times. And here addressing just pricing concerns. So Joshua mentions the pricing structure, but including a brief ROI example over testimonials could further justify the cost and alleviate concerns. This is what I would look at.

Now, I would not take every single suggestion it comes back with, but I have yet to write a reply where I didn't then change the email before I actually sent it out. And so it is fundamentally improved every single important sales email I send out. I'm not saying I'm using this for every single email that I write, but if it's something that might, it's like whatever, 10, 20, $30,000 type of contract, I'm like, yeah, it's probably worth another minute of me going back and forth.

No, I don't let it write my email. So I, in this case, so I don't think they're good enough yet to blindly send it out. I'm getting close to it and I might write a flow where I kind of like, I get it to draft the email and then I just hit send. Um, but right now I'm more in the flow where I just get it to improve the one that I send.

Privacy Concerns and Data Usage Policies

Do you worry about sending personal to the company or do you think I mean, you can, so the privacy policy of OpenAI is very clear. None of this information is getting trained on in any way, so long as you're on the commercial version. Their privacy policy is pretty robust, saying this information doesn't go anywhere other than on your account.

Now, you can, there's a whole bunch of conspiracy theories that, yes, they say one thing, they do another, but in that case, I might as well also question Microsoft, Google, any single thing that we're using on the internet. So... Yeah.

So the moment you're on chat duty plus, the data is not used to train the models. You can actually, like, literally in your account, you can actually see that, by the way. I think it says it even in the, yeah. So in the data control here, there you go.

You've got a chat on, so it is even. No. That is the history. That's very different.

That means that it remembers it for my use. As you can see here, chat chaining. This works with private and opted out of training.

I'm using Teams, but that's the same thing when you're on the ChatGPT Plus account. Sorry? It's not default. Okay, the default, I don't know. Okay, I don't know. I'm going to stop there.

Conclusion and Q&A Session

Obviously, I'll be here for a while for questions, but hopefully this was useful and you can build your own GPTs from now on.

Cheers. Thank you.