Generative AI - Deep Dive

By Wayne Brown

 This is a draft - made "public" for my very few followers 


THE NATURE OF GENERATIVE AI

In what sense does it "understand". It's easy to fake understanding if AI is presented with an example of something that's all over its training data. But pick an original photograph. Much less impressive.

The original photo - my mother-in-law (104) 
A mutual love affair

PERPLEXITY AI DESCRIBES THE PICTURE 

Image shows an elderly woman sitting comfortably on a couch or chair. She appears to be wearing a black sweater or top and has short gray hair. Next to her is a fluffy white dog, likely a poodle or similar breed, sitting close and providing companionship. The woman seems relaxed, holding some papers or documents in her lap, perhaps reading them. The setting looks cozy with cushions and a wooden or laminate floor visible in the background.

It's not too terrible, actually. Most humans would not identify the dog as a "Golden Doodle." Humans may not mistake the coffee table for a floor, but most would miss it entirely. Placing it in the "background" is typical of AI's problems with spatial arrangements, subtly thinking that the old lady is reading the documents rather than cuddling the dog.

Most impressive is how AI captures the mood.



 A 104 year old woman reclining comfortably on a couch. She appears to be wearing a black sweater or top and has short gray hair. Next to her is a fluffy white golden doodle,. The setting looks cozy with cushions and a wooden or laminate floor visible in the background
(slightly edited to age the lady and identify the dog without much success - not a Golden Doodle cut)
(this is Microsoft DALL-E)

"Night Cafe nails it, getting the dog breed right and picking up lady's age and the cuddle theme. No nonsense about the "floor" or papers.

Another Night Cafe shot picks up the classic Doodle face

"Playground.ai"
Nails the mood, gets rid of irrelevant detail

These are examples of AI at its best. A bit of tweaking helps with image interpretation. AI has no trouble coming up with cute pictures on the theme. Looking "under the hood," we see that there are a lot of pictures with that theme, making it easy for AI to come up with something "similar."




WHERE DOES THE "ART" COME FROM

Same prompt, multiple results. Often amazing but no way the "promoter" can take any kind of credit. Provide a few examples of AI images from given prompt. Tremendoius differences and epic fails point out the tenuous connection with the prompt. The result is amazing but in some way it's just a pretty good creation based on a known set of images.

There is a statistical concept that's applicable. Any measurement must be consistent before you wonder if anything real is being measured.


Comments

Post a Comment

Popular posts from this blog

Facebook and Bing - A Killer Combination

By Wayne Brown
Image
 This is a partial list of groups I belong to that focus on Generative AI art: Facebook creates community. I use BING as a kind of Generative AI university: BING U.  I prefer the role of student, so my job is to find good teachers. Think of me as the guy who shows you around the campus and knows his way around, but he's a student himself. Microsoft BING is free and very powerful. It's an excellent way to learn Generative AI and a standard tool that everyone in my imaginary school has. The way to find a teacher is to see their work in a Facebook group. Then, save the picture that caught your eye in a Word document (or document editor of your choice). For example ... Joe McIntosh What we have found here is not only the prompt for the picture but a teacher who shares the prompts for all his pictures. Look for artists who work primarily with BING. Don't get distracted by other platforms or techniques. Learn to do one thing well before shopping for better tools. The best Facebo
Read more

A Process ...

By Wayne Brown
Image
She will be 105 in September. She struggles to recognize her four children, let alone her countless grandchildren. Today, she said, "Nobody told me how to die." She has always wanted to die in her own home, an oversized place that has always had room for an endless parade of visitors. She has always been very "social." Although she struggles to remember the old friends dropping by, she chats happily "on automatic." There are a lot of relatives. She was raised on a farm near North Battleford, Saskatchewan, where her Russian parents raised 16 children. I asked "AI" - specifically Claude 3 Sonnet , to describe this situation in a short essay. I have taken the liberty of letting Grammarly , my AI writing assistant, make a few corrections to Sonnet's purple prose. This is the slightly edited version: She sits in  the twilight of her extraordinary life, a living embodiment of resilience and grace. At 105 years old, her eyes, dimmed by time, still h
Read more

Warp Speed Generative AI

By Wayne Brown
Image
Jump to Warp Speed Warp 2 Warp 3 Warp 4 "Warp Speed"  Illustration of " Language is a Virus " Laurie Anderson Completed in the time it takes to perform the song For a very long time, I have tried to master information processing tools and "wire" them together into a "machine" that suits my purpose. "Warp Speed Generative AI" combines several powerful families of AI tools to turn an idea into one or more pictures. Along the way, we get an explanation of the original idea, references and a detailed explanation of the style. I call it "Warp speed" because AI seems to give me whiplash as it takes my idea and runs with it. I just hang on and enjoy the ride. "Warp Speed" is a response to the current obsession with prompts: how to write them, when to share them, how to monetize them, what books to read, what courses to take and so forth. AI can write pretty good prompts. "Prompt Engineering" is an endangered speci
Read more