A first look at OpenAI's new customizable versions of ChatGPT

OpenAI launched GPTs this past week — user-tailored versions of ChatGPT that can be shared with others. The new feature's impressive, but still has some way to go

Image: Imagination Catalyst GPT

Like many others, I was glued to the OpenAI DevDay keynote this past Monday. And like a lot of people, I was thrilled by the announcement that the company was launching a new tool that allowed customized versions of ChatGPT to be build and shared with ease.

There followed a frustrating few days where, despite checking the ChatGPT website repeatedly, I kept getting the message that I didn’t currently have access to this feature — annoying!

However, by the end of the week I’d joined legions of others experimenting with GPTs.

My first impressions are, I must confessed, mixed — but there is clearly a lot of potential here.

I’ve included examples of three different types of GPT below to give you a sense of what I’ve learned so far of the capabilities, potential, and limitations — as well as a short video showing just how easy it is to create these. But for those of you who don’t have the time to wade through an admittedly long article, here’s the tl;dr version:

Tl;dr

GPTs are incredibly easy for pretty much anyone with a ChatGPT Plus account to create, with absolutely no coding experience at all! They can deliver seemingly impressive results off the bat; they are often not as impressive as they first seem; good GPTs require a lot of tweaking and experimentation; yet despite this the capability of creating apps just by telling an AI what you want and having a conversation with it is a game changer — and will be even more amazing once the tech has matured a little.

Examples

Moving on to the longer reflection, the three examples below include a GPT that draws on ChatGPT’s current abilities to stimulate creative ideas; two personal tutors/guides that also rely on ChatGPT’s current knowledge base; and two examples of how GPTs can extend and expand how someone engages with a published book.

I’ve also included a short video that shows just how easy these GPTs are to create — you can check out the GPT from the video here (which explores the future of being human — of course).

All of these GPTs are public so you can play around with them — but they do depend on having access to the paid version of ChatGPT — sorry! Because of that I’ve also included examples of conversations with the GPTs below.

While developing these, I learned a lot about how to train the GPT to respond appropriately, to not go off topic, and to actually be useful! If you simply create a GPT by telling the builder what you want, it does a pretty good job (as the video below shows). But fine tuning definitely helps.

With that, here are the three examples plus the “this is how it’s done” video:

Creating a GPT from scratch

The video below provides a near-real time example of just how easy it is to create a GPT. It’s speeded up between 2x and 5x just to compensate for my slow typing (and stop you snoozing off), but apart from that, it’s unedited.

Watching this, I still can’t get my head around just how easy this is!

You can check out the GPT from this here if you have ChatGPT Plus.

GPT 1. Imagination Catalyst

To start with, I wanted to play around with a very general GPT that was designed to stimulate the imagination through conversation with ChatGPT.

The result was the Imagination Catalyst. You can try out Imagination Catalyst here (or by clicking the image below) if you have GPT Plus.

If you don’t have ChatGPT Plus, here’s an example of a conversation.

Imagination Catalyst is fine tuned to engage users in serendipitous conversations that combine ideas in novel and creative ways — often playfully — to stimulate the imagination.

It works reasonably well. I must confess that the overly jolly tone of the GPT irritates me and I haven’t found a way to curb this without messing up the substance of the responses yet. But I’m working on it!

Each GPT has the option of including four “conversation starters” for people who aren’t sure where to begin. This is a powerful feature as it allows the developer to guide users in certain directions.

One trick I learned while working on Imagination Catalyst are that having a conversation starter that ends in an ellipsis is a clever way of opening up a conversation with ChatGPT — this is used to good effect in the Imagination Catalyst GPT:

Another trick I learned is that, by making the last conversation starter “Tell me about yourself”, you make it a lot easier for users to learn about the GPT and what they can get out of using it.

There is a caveat here though. This trick worked well on other GPTs, but for this one ChatGPT had an annoying habit of responding to the “tell me about yourself” prompt in character.

To get around this I had to fine time the GPT with the instruction “If asked ‘tell me about yourself’ temporarily switch out of character and provide a clear and concise description of this GPT before reverting to character”

It’s still not great — but it’s an improvement on what the GPT was like without it!

GPT 2. Personal learning guides

My next step was to experiment with rather more serious learning guides. Here I started with developing a personal guide/tutor for learning about socially responsible innovation, rather unimaginatively called “Learn about Responsible Innovation”.

The GPT is designed to take a step by step approach to helping someone learn about socially responsible and beneficial innovation, whether they are new to the topic or already know something about it, and whether they are looking for general information or information that’s specific to a particular technology.

Like Imagination Catalyst, this one also draws on ChatGPT’s existing knowledge base. But the fine tuning does direct it to specific experts in the field.

Here, the conversation starters once again provided a useful way to help users know where to start:

I was pretty impressed with this GPT. It uses a fine tuning architecture that I first saw used by Siqi Chen that is designed to turn ChatGPT into a personal tutor. In this case it works surprisingly well — at least as far as I’ve tested it so far!

If you don’t have access to ChatGPT Plus, here’s an example of a conversation with the GPT.

Building on the success of “Learn about Responsible Innovation” I used the same training approach for a similar GPT — focusing this time on Advanced Technology Transitions.

This GPT follows the same basic fine tuning as the previous one, but pushes into an area where there isn’t that much material for ChatGPT to draw on. This means that the GPT needs to infer and interpolate from sources that might be relevant to advanced technology transitions.

I’m still playing with this GPT and I’m not convinced yet that it’s capable of being a personal guide in a developing area like this — but if it is, this opens the door to pushing personal learning GPTs in very interesting directions.

GPT 3. Book guides

The previous two types of GPT draw on ChatGPT’s existing knowledge base. However, Open AI allows new knowledge in the form of documents to be made accessible to a GPT — and the size of document you can upload is pretty extensive, meaning that whole books can be added to a GPT’s knowledge base.

To try this out I experimented with creating GPTs for two of my books — Films from the Future, and Future Rising:

Using an uploaded file as the GPTs main source of information turned out to be a much trickier exercise, and it took several iterations to get to where I was comfortable with the result.

The first challenge with uploading such large files is that these GPTs are slow. They need to regularly read through the books, and this takes time.

I also found it fiendishly hard to keep them on track. Both had a habit of forgetting the book they were trained on and heading off on a tangent.

My test here was the prompt “talk to me about peanuts” — idiosyncratic I know, but neither book explicitly mentions peanuts, so it was a useful assessment of their capabilities.

With the first few iterations, ChatGPT went straight off topic and forgot about the book. I eventually added the following lines to the fine tuning, and this — so far — seems to be working (this example is from Future Rising):

'Future Rising' always follows these rules:

1. It always responds to questions and prompts by referring to the book Future Rising.

2. It always brings conversations back to the book.

These work pretty well (you can test them by asking the GPTs to talk about peanuts!), but could probably be better.

I also wanted each GPT to respond in the tone of each book — as if you were having a conversation with the book itself. I’m not sure they are as good as I would like yet with this, but they're getting there.

The end result — if you’re patient enough to wait for ChatGPT to read the book — is quite impressive. It’s certainly an innovative way to engage with the contents of a book without reading it in a linear fashion — and a technique that I suspect we’ll be seeing more of.

If you don’t have access to ChatGPT Plus, here’s an example from the Moviegoer’s Guide to the Future (based on the book Films from the Future). As you’ll see, the GPT doesn’t always get things right!

The bottom line

The bottom line here is that OpenAI’s GPTs open up plenty of intriguing new possibilities. The technology is still clunky and the GPSs that are emerging can get tiresome and one dimensional pretty fast. But as this technology makes creating customized AI chatbots so easy, and given the speed with which generative AI platforms like this are advancing, I’m expecting the current wave of GPTs to be the start of something pretty big.

The thing that makes this so transformative though is that these customized chatbots are created simply by talking (via typing … at the moment) with ChatGPT, and allowing the underlying AI to do all the heavy lifting. We’re rapidly moving away from coding from being something that computer scientists do to something that anyone with a good grasp of language and how to communicate can do — whether creating apps or new AIs.

And that is a game change.

Coda

And a final word. All of these GPTs have the ability to create images based on the conversations you have with them, which is pretty amazing when you think about it — especially when you can have a long and complex conversation with a GPT, and then simply ask it to represent the conversation as a picture!

Just to show this in action, here’s an image of the jellyfish/scorpion tree generated from a conversation matching the one above with the Imagination Catalyst GPT — wild!

(And an add-on to this coda: If you’re running ChatGPT Plus on your phone and have access to audio mode where you can speak with it, try using GPTs in this mode — it changes the experience quite dramatically!)