This week's short post is for anyone who has only had a bit of a play with chatGPT, not the experts. This is beginners only and it may seem like it is bordering on science fiction.

First things first. There is a tremendous amount of hype surrounding AI. Tonnes of it. Usually coming from people who have a vested interest in selling you something, typically a subscription of some nature. Amongst all this shite though, there is some useful stuff, and creating your own chatGPT is one of these I reckon.

What is cool is that creating your very own ChatGPT is far easier than you might think. I am going to show you how to do it in this short post.

The first question though is why bother? Why not just use chatGPT as it is? Well, part of the reason is that we use it really badly. We are crap at creating prompts and asking questions, this means the results we get back tend to get less than satisfactory answers. Being humans we think chatGPT or any LLM for that matter, is not as good as people say, after all, we cannot be to blame, surely! This is a real shame because, with a little bit of effort we can get a tremendous boost in the quality of the answers we get from chatGPT, or any large language model (LLM).

Often, the first step is the hardest when trying something new, and this is true here when trying to create a chatGPT. The first thing you want to do is to have some idea of what you're trying to build, your "vision". This is far harder than actually building the custom chatGPT. What do you want your AI companion to achieve? Knowing "what" you want to build is often the show stopper simply because you have never built one before.

Is it to write sales pitch emails for you? or to generate Python code to solve a user problem? Are you already shrugging at this point?

One way to think about chatGPT is like a massive library. If you went into the world's largest library and asked for a book on general relativity, you could probably find hundreds, if not thousands of books on the subject. Which should they bring you? Chances are they will give you a basic, popular book, which may or may not be what you are after. If you go into the same library and tell them that you are a physics professor and you are looking for books to put together a series of lectures aimed at postgraduate physics students who are well-versed in the subject, they will bring you a different set of books. If you explain further that you are doing a historical review of the subject, they may bring less technical books and so on.

Creating your own GPT allows you to tell the chatGPT to be more specific. What is great though, in our previous example, if I then ask for information on say quantum mechanics, they will bring me historical books on quantum mechanics aimed at postgrad-level physics students. If I say "thermodynamics" I will get historical books on thermodynamics. I don't need to keep telling it that I am a "physics professor and you are looking for books...".

In other words by giving better instructions of what we are after the assistant can bring back more relevant material. This is one way of thinking about a custom chatGPT. What many don't realise is that chatGPT and LLMs already do this to some degree. They have their own set of prompts behind the scenes that are applied when you type a prompt.

Ok, this may be of some help, but sometimes, it is easier if you have a look at someone else's examples. The old "here is one I made earlier" for those brought up on 80s television. Do such things exist? Of course.

The first thing you can do is take a look at the many custom GPTs that have been created already. There are tens of thousands of them. The idea is that it is just like the app store, you can find an "app for that". Many of these prebuilt GPTs will have some paid feature, but you can use them for testing, getting ideas or seeing how things are done.

You can find them by clicking on the "Explore GPTs" link on the left-hand side of your chatGPT web interface. You will usually be presented with "Featured" GPTs and "Trending" GPT, "By ChatGPT", "Dall-E", "Writing", "Productivity" and many others. Have a play and try some of these.

There are lots of GPTs out there, spend some time having a play with these, this is time well spent. If you type "email" into the search bar, you will get a dozen or more "Email assistants" to try. These are all designed to write better emails. Type in "Python" and you will get a bunch of copilots and code generators to choose from. You can also find image generators or access to tools you may also be using such as Canva.

Don't get me wrong, there is also a tonne of shit out there trying to get you to sign up to some Saas product that will disappear in the next 6 months. Playing with all these examples, good and bad, will give you some ideas, that is the real point. Then go back into the "Explore GPTs" and look up into the top right hand corner.

You will see the green "+ Create" button, this is the button you use to start your bespoke chatGPT journey. Click on that fella and we are on our way to creating our first New GPT.

When you do you are presented with a couple of text boxes. Give your GPT a name, and a description. The next box is important, the "Instructions" box, this is where you tell it what it is. Don't worry, you can come back and tweak this, you don't have to get it right the first time. In my example earlier, the physics example, I might put something like

"You are a physics professor and you are creating a series of lectures aimed at university physics students. The user prompt will give you information on the subject and the year of the student, for example, "quantum mechanics 1st year", "Thermodynamics 2nd year", "electro magnetism 3rd", "Quantum field theory post graduate". You will create titles and contents for a 10 lecture series. Each lecture will run for approximately 1 hour."

when I test this all I have to enter into the prompt is something like "quantum mechanics 2nd", "superconductivity post grad" and it will generate 10 titles and some description of each lecture. For the prompt "superconductivity post grad" it returns

*"*For a postgraduate course on superconductivity, the lecture series should delve into both the theoretical foundations and cutting-edge research in the field. This advanced series will cover the phenomenology, microscopic theory, and applications of superconductivity, along with discussions on high-temperature superconductors and quantum technologies. Here's a proposed structure for your series:

Lecture 1: Introduction to Superconductivity Historical overview and fundamental phenomena (zero resistance, Meissner effect). Types of superconductors (Type I and II). Experimental discovery and significance in physics.

Lecture 2: Phenomenological Theories...*"*

and so on. When you are happy with your instructions you "Save" your GPT, you get options to save for you, anyone with a link (so you can share with friends/colleagues) or everyone. You are then good to go. That is it! You have created your own GPT. Name, description, and instructions, that is all you need. I said it was like science fiction eh?

I am going to wrap this up in a moment, but one final thing. In these posts I am talking about chatGPT, this is because it is the most accessible LLM at the moment. Anyone can use it, even those without any technical expertise whatsoever. The idea of a custom GPT and what it can do for you is true of any LLM out there.

One final thing to ponder, it is not a question of when will AI get here, it is already here and people need to start getting familiar with it. Learning to write prompts and creating bespoke GPTs is part of that process.

In the next post we are going to talk about the open AI assistant, this one is a bit more techie, but not much more to be fair.

Bryan
ps. this is written by a human, me, if it is shite that is on me, not the AI.