Weekly Roundup - Tue, Jul 4 2023

Two news items landed this week worth mentioning. 1. Midjourney announced a new prompt parameter. And 2. GPT-4 may not be as technically sophisticated as previously thought.

Weekly Roundup - Tue, Jul 4 2023
by Matt / Midjourney: dancing sunset --weird 3000 --s 1000 --seed 3357139127

Let’s get --weird

On the heels of announcing model version 5.2, Midjourney has announced a new parameter: --weird. No, literally. The name of the new parameter is --weird. The --weird parameter lets users explore unconventional aesthetics and create unique and unexpected image outcomes. It accepts values from 0 to 3000 and is highly experimental. If you want both conventionally attractive and simultaneously weird images, try combining higher --stylize values with --weird. See the Midjourney documentation below for details.

Midjourney Weird Parameter
Explore unconventional aesthetics with the experimental `--weird` parameter. This parameter creates unique and unexpected outcomes.

I look forward to experimenting with the new parameter in our budding wallpaper collection. Pastel sunset wallpapers are coming next week.

GPT-4: Eight GPT-3 Kids in a Trenchcoat

GPT-4 is like that classic trop of multiple kids stacked on each other, pretending to be an adult and getting into the cinema. GPT-4 has likely been eight instances of GPT-3 all along!

Previously it was rumored that GPT-4 was a single model with over a whopping trillion parameters. It is believed (though not confirmed) that GPT-4 isn’t a single monolithic dense model (like GPT-3 and GPT-3.5) but a mixture of 8 x 220-billion-parameter models.

What’s the implication? It means other closed and open-source platforms aren’t as far behind as initially thought. This is good news. Regarding prompt making, I have noticed result variance across paid accounts. Does that have to do with different accounts being assigned a different domain expert? More to study regarding the Mixture of Experts (MoE) model GPT-4 likely employs. Craft best-in-class prompts mean being able to reproduce the results across accounts.

Key takeaway - GPT-4: not groundbreaking, but still the best language model available.

GPT-4 is 1.76 trillion parameters in size and relies on 30 year old technology
OpenAI GPT-4 is said to be based on the Mixture of Experts architecture and has 1.76 trillion parameters.

Lastly, here are the rest of the "weird" photos I generated in Midjourney. To me, they didn't really turn out that weird at all.