I've gently dipped my toe in, and here's a quick review.
Surprisingly good:
Composing lyrics, poetry or short stories.
There is a recognisable style about the prose, which hints that you're dealing with AI.
That's less present in the songs and poems; probably because the structure of the form forces variations.
Terrifying:
Comprehension of language and composing pictures.
For comprehension, I took a manga series that #3 son reads, fed the AI character profiles for a fan wiki and asked it to predict events.
Most predictions were a hit in the sense that they've happened in the story, or events are lined up to make them happen.
What really amazed me was the lack of prompts for these events in the character profiles (There's a brief section describing friendships and rivalries of each character, but that's all).
I think this illustrated AI's much vaunted potential to spot minor patterns that people can't see - one where it excels at medical scan analysis.
Pictures, I know it's all about overlaying billions of sampled photos, but the results are remarkable.
A little too smooth and rounded to pass as fine art or photographs.
Facial expressions are currently bland, or direct samples.
You won't be getting Munch, Toyokuni or Da Vinci here.
Do expect deepfakes of Keir Starmer eating a bacon sandwich - very badly, or praying to Mecca while waving a Hamas flag.
A bit crap
Dialogue, with a human.
It's worth reviewing the background
https://en.wikipedia.org/wiki/Natural_l ... processing.
Weizenbaum's Eliza, has been around for almost 60 years, was obviously a robot, but could conduct one end of a dialogue.
We now have chatbots embedded into online commercial apps including banking and shopping.
I get the feeling they do a good job at reducing the supplier's call volume.
I find they end up directing me to a Frequently asked question list (which I've already read) and can't work "outside that box".
I had a go at a system which conducted raw conversations.
Its ability to comprehend my words was remarkable.
The responses usually started extremely well, but tailed off into chains of predictable platitudes.
"It is through such things we grow and learn" - profound once, irritating in every second response.
The conversation engine really fell apart when pushed above a 1:1 dialogue.
A second computer character remained mostly quiet or echoed the words of the first.
(Which reminded me of an elderly pair of twin spinsters who lived next door to my grandparents when I was young, one was the talker, the other the echo).
With two human characters, the AI struggles to follow the sense of the conversation.
The other weakness appeared to be remembering facts form the conversation.
The AI was extremely good at inferring and remembering emotions (likes and dislikes), it had an immense background knowledge (wasn't wrong-footed when I suggested shopping for a Durian), but it completely forgot that we have driven to the shops by car.
So there's a lot of really powerful stuff there, with potential for good and evil.
Also a few areas of weakness, which a knowledgeable "blade runner" might exploit to distinguish man form machine.
