Misadventure in configuring my own GPT

Or how an adult man spent a week talking to AI about cats. The Catventure Creator

When OpenAI announced the option to create your own, purpose-specific GPT I was quite excited! Generative Pre-trained Transformers (GPTs), especially GPT-4, have shown their incredible ability to provide many valuable services. Generating code or creating a brief summary of a report,? conducting SEO or acting as tutor the list of applications is huge and ever growing. The cost to that is “compute” – this term describes the computational cost of running the GPT on a dataset. ChatGPT is a very generalist model, and the datasets are both diverse and very large. The compute cost is equally as impressive. A narrow-scope GPT focused on a relevant task can be both very useful and yet have a lower compute cost! Glorious! I should definitely build one, I thought to myself…

?????????????? There I was looking at the screen, the configuration font on the left, and the test prompt on the right. But what would my GPT be focused on? There is plenty of clerical work to be done alongside the kind of Biomedical R&D that I am doing: requirement engineering, specifications, testing protocols and other quality related stuff. That’s the kind of work a GPT would be good at. This, however, would make for a very boring demonstration, not to mention that I would either have to come up with a reasonable test case or skirt the edges of the NDA I signed at work. I wanted to keep it simple. Really simple. And so I thought: a popular opinion is that Internet is for pictures of cats. But hyper-realistic cats generated via Dall-E have that unnerving “uncanny valley” feel to them. So they should be cutesy and stylized… Stylized to DnD adventurers! DnD is more popular now than it has ever been – Critical role is topping Twitch charts. But it has to be something more. Maybe some utility to the picture? Let’s make this a calendar! And so the Catventurer Calendar Creator was born. The creator does this:

·??????? Takes a picture as input

·??????? Creates a highly stylized DnD themed cat-adventurer, attempting to maintain some features: arm position, head tilt or turn, glasses etc.

·??????? Takes month and yeas as text input and generates a calendar grid

·??????? Combines the two and outputs a monthly calendar

Well, what did I get out of it? Was it a success? Kind of… Some outputs are absolutely grand!

Excellent! What’s the catch? The catch is… The Catventurer Calendar Creator only outputs something like that 10% of the time. The rest of the time you need additional prompts and correction, and time to time nothing happens at all, as the GPT just returns and error… And there are more “esoteric” failures as well!

Failures range from poor placement of calendar components, to generating an illegible “calendar” via the Dall-E interface, to doubling up on the Catventurer to something else entirely! I don’t know the creepy guy in the centre bottom picture, and I don’t want to know who that is. And that is all not to mention the endless “Error Analysing” messages than mean the GPT is restarting the processing. It will give up after 3-5 errors.

?????????????? I have attempted every debugging and fool proofing measure I could: I tried specifying the dimensions in pixel (both for the Dall-E output and relative positions), asking for error checks at every stage, uploading my own code for calendar generation, asking for thorough error reports… And the only thing that actually gave some results is making sure the merging operation between the calendar and the Catventurer is restarted when a bad result is achieved instead of trying to fix it. The fixing would actually confuse the GPT further.

?????????????? I went ahead and made my own calendar routine, even tried to create my own calendar creator… The script makes a call to GPT-4V to get a description of the photo, sends the description to Dall-E and merges it with a calendar I created beforehand. It works every time, and yet the GPT I instructed in the browser does not…

?And then I realised… The functionality of creating your own GPT is in early development stages, and OpenAI aren’t developing AI + wrappers, they are looking for AGI. And perhaps it’s what is supposed to join the parts of the calendar is what is going through development, and I can hardly change that with the prompting. And so content with that thought, I decided to check out some GPTs openAi themselves have created. I clicked Cosmic Dreams and one of the suggested prompts.

And therein lies a lesson... Don't overanalyse things. Still, there are good takeaways:

• Make your prompts very clear
• Remind the GPT to make reality checks
• Use objective input, i.e. pixels
• Restarting is better than fixes, especially on a short workflow like this

I see incredibly potential with this type of GPT: calendars, organisers, customized board games... The list goes on! I intend to continue experimenting with GPTs.

I plan to post Part 2 of Dall-E vs Proverbs next time and hopefully a post on my AR project the week after. In closing - here's some good outputs I got.

Give the Catventure Creator a try and tell me how it goes for you!