The basic assumption I make when I form an opinion about new things (technologies, processes, startups, etc.) is that my opinion will have to change over time.

As the amount of knowledge one has increases, the first opinion becomes less relevant.

The last weekend I participated in hackathon at Surf Incubator. The focus of the hackathon was AI, or more in particular OpenAI APIs. Every team had 3-5 hours to hack a new AI app.

The one I chose to build was FashionGPT, which would take your selfie or photo and give you 3-5 options on how you would look with the latest fashion outfits from luxury brands such as Gucci, Giorgio Armani, etc.

It is a virtual trial room. You dont need to try new outfits, just see what you’d look like so you can share on Social media and get feedback before you decide to buy.

I worked with one other person, who quickly gave up and I complete the 3 modules – Javascript + HTML front end, Python at the back with Django on AWS. I pushed it to Github.

All the version 1 did was take your photo, then I asked ChatGPT for names of 100 top luxury brands, 100 top colors and 100 top outfits styles in 4 seasons (Fall, Winter, Spring, Summer).

The “prompt” to ChatGPT was a random combination of brand + color + outfit, using the create image endpoint, and then variations endpoint for 3 different colors and styles.

Until yesterday I had a very poor impression of the term “prompt engineer”. I thought that was a BS job. Since there’s no specialized “google searcher” role, why do you need a “prompt engineer”.

Turns out teasing information from LLM is a lot harder when you are not expressive. Most people are not.

I spend nearly 1 hour trying to build a random prompt base so the model could give meaningful new images from the existing set of images plus the “secret sauce” – the prompt.

“Prompt engineer” is a thing. I was wrong.