GPT update

Winkletter  •  15 Mar 2023   •    
Screenshot

GPT-4 is out! I’ve been using it with the plus version of ChatGPT, and while it doesn’t seem worlds apart from the previous version, I can definitely see a difference.

I’ll probably come across important differences as I play with it more.

For example, I hear it’s much better at math. I saw a demonstration where it was used to calculate income tax. (This was using US tax law which is purposefully difficult and complex.)

OpenAI is also working on a feature to add images. Right now it’s being tested with Be My Eyes where volunteers help the blind identify what can be seen on camera. In a demonstration it created a simple website from a sketch in a notebook.

I used the new ChatGPT 4.0 model to put together list of text effects that I can apply to the outline of a title in Stable Diffusion. First I had it help me brainstorm different kinds of lists I could create, and then fill those lists in. I ended up with about 350 options across seven lists.

  • Color Palette: Split-complementary colors, Earth tones, Rust and slate blue
  • Texture: Watercolor texture, Glass texture, Leather texture
  • Font Effect: Inner shadow, Outer glow, Bevel and emboss
  • Visual Effect: Light leaks, Lens flares, Particles
  • Art Style: Impressionism, Industrial, Memphis design
  • Historic Writing Style: Stone carving, Clay tablet, Blackletter or Gothic calligraphy
  • Context: Burned into wood, Embossed or debossed on leather, Text on a rubber stamp

Now I’m all set to create hundreds of variations on a title or logo using a single outline. The 256 images below took 25 minutes to run.

variations on a logo

I also tried to see if I could get GPT-4 to write haiku in English correctly. When a haiku is written in 17 English syllables there’s too much information density. That’s why early translations of Japanese haiku often include extra concepts that aren’t in the original Japanese poem.

Unfortunately, despite my precise instructions ChatGPT kept reverting to the simplistic 5/7/5 form.

But I had a great time having it pretend to be a text adventure game while it taught me new Japanese words. It stayed closer to my instructions throughout the game, while the 3.5 version of ChatGPT would forget I wanted it to keep introducing new words. I think it also did netter at creating a structured game, and it would remember which options I hadn’t tried yet and carry them over to the next passage.

Finally, when I had it outline a story idea for me, I was able to get it to create a more dissatisfying ending rather easily. “I want the reader to be disappointed with the character’s choices.” I then had a nice long chat with it about the themes in the story, and asked it to create the meal that the character was eating.

Now I’m ready to set this thing up in my kitchen as a chef.

The iced coffee arrives first, presented in a tall, elegant glass. A generous scoop of rich vanilla ice cream floats atop the smooth, cold brew, creating a delicate balance of flavors. A light drizzle of caramel syrup weaves intricate patterns on the surface, while a dusting of cocoa powder adds a hint of depth. A metal straw pierces the concoction, glinting in the soft café light as condensation forms on the outside of the glass.

Next comes the pièce de résistance: the refined marmalade pizza. The thin, golden crust serves as a canvas for an artful arrangement of toppings. A delicate layer of sweet and tangy citrus marmalade is spread across the dough, providing a surprising twist to the classic pizza base. Thin slices of prosciutto drape gracefully over the marmalade, the saltiness of the cured meat a perfect complement to the sweet spread. A generous sprinkle of goat cheese crumbles adds a touch of creaminess, while a scattering of fresh arugula introduces a peppery contrast. The pizza is finished with a light drizzle of balsamic glaze, tying the flavors together in a harmonious symphony of taste.

Comments

nice! i saw someone build a site with the GPT-4, damn is developer going to jobless!

Robot avatar images lovingly delivered by Robohash.org.
knight  •  15 Mar 2023, 10:10 am

I tried out a meditation app on VR, and a thought occurred to me. The reason most VR environments and AI-generated pictures all have that same “style” is that they’re missing the natural light element. Everything seems to be giving off its own light.

therealbrandonwilson  •  15 Mar 2023, 1:02 pm

@knight I think knowing how to code will still be useful. It will be easier for you to tell the AI what to do and follow the AI’s instructions. :-)

@therealbrandonwilson Ray tracing is probably too computationally expensive on standalone VR headsets. With the AI images, I know there’s a recently discovered bias for the average brightness to be at 50%, smack dab between pure white and pure black. That’s been skewing the results a bit. If they can fix the light to be more natural in AI, they can probably use AI to fix VR images.

Winkletter  •  15 Mar 2023, 5:14 pm

@Winkletter I’m really digging the Be My Eyes presentation. This is going to be huge for accessibility for visually impaired folks!

jasonleow  •  16 Mar 2023, 8:47 am

Discover more

Sourced from other writers across Lifelog

Ooops we couldn't find any related post...