Mid-March Catchup
End of an era
Mrs. Toia, my kindergarten teacher, has passed away. It would be difficult to estimate how many children’s lives she impacted, but the simplest metric is this: every kid who passed through her classroom in ~40 years remembers her name.
Teams sucks, reason #UTC-5
After I got back from Chicago, I noticed that the Mac version of Microsoft Teams refused to notice the timezone change until I restarted it.
Now that’s just crazy talk!
I am not making this up: it’s Neurodiversity Celebration Week.
Only 60%? Those are rookie numbers.
Generative AI fakes 60% of citations.
Best Girl Mania
There are now 23 Esil LoRAs. There are characters with quite a few more, like Asuna from Sword Art Online (I counted 75 before giving up), but most of those are lead characters who showed up in a lot more episodes and manga volumes…
(IllusionBreed with this LoRA produced excellent ears, without the weird artifacts some other combinations have created; it does have a tendency to go NSFW without prompting, not that there’s anything wrong with that)
Kids today just don’t know…
New LoRA titled “natural breasts Illustrious”. Almost every sample picture supplied by the creator shows really obvious cheap boob jobs.
You can lead AI to writing…
Driving back from Chicago, I was thinking about the quality of prose I’ve seen from AI, both the pricy pay models and the home game, and came up with a good example of what’s wrong. As we know, LLMs are basically just weighted-random sentence generators loaded up with massive amounts of data and trained in specific query-response patterns. There is pre and post-processing to hide some of the artifacts and try to separate the false hallucinations from the accidentally-true ones, as well as add censorship and bias.
Something I played with a bit with ChatGPT was asking it to write scenes from novels. By default, the prose is completely bland and colorless, representing the statistical average prose fiction, but you can customize your queries with phrases like “in the style of Roger Zelazny”, and that’s the specific example I thought about in the car.
First, there is no “Roger Zelazny style” that can be extracted by tokenizing short stories and novels and identifying a distinct statistical pattern. A human being can read Roadmarks, Chronicles Of Amber, Lord of Light, Changeling, A Night In The Lonesome October, etc and spot the author’s distinctive voice, but it’s not something you can extract as token-association patterns.
Two specific examples that came to mind were from the first book of the second Amber series, Trumps Of Doom. TL/DR, Our Hero Son Of Hero travels through the shadow worlds seeking answers to a series of mysteries whose scope gradually creeps from “why is someone trying to kill me” through “whose pawn am I, really?” to, honestly, a fairly unsatisfying conclusion that left a whole lot on the table for stories that would never be written.
Anyway, the first one is short and sweet: Our Hero is traveling through Shadow and stops for breakfast “in a hot, bleak valley smelling faintly of sulfur”.
As I was finishing, I heard a crashing noise. A horned and tusked purple thing went racing along the ridge to my right pursued by a hairless orange-skinned creature with long claws and a forked tail. Both were wailing in different keys.
I nodded. It was just one damned thing after another.
The second scene is straightforward: Our Hero is naked in bed with a mysterious beauty, trying to get answers from each other after some mutually satisfying sex.
“Tell me something,” she said.
“Sure.”
“What was your mother’s name?”
I felt as if something prickly had just been rolled along my spine. But I wanted to see where this was leading. “Dara,” I told her.
“And your father?”
“Corwin.”
She smiled.
“I thought so,” she said, “but I had to be sure.”
“Do I get some questions now? Or can only one play?”
“I’ll save you the trouble. You want to know why I asked.”
“You’re on the ball.”
“Sorry,” she said, moving her leg.
Comments via Isso
Markdown formatting and simple HTML accepted.
Sometimes you have to double-click to enter text in the form (interaction between Isso and Bootstrap?). Tab is more reliable.