Doors, edumacation, gratuitous AI, and *useful* AI

Doors, edumacation, gratuitous AI, and useful AI

Thu 11/27/25 11:56pm

Comments: 0

Fun

Door, More Door, Gone Door

(insert meme here)

The back door won’t be installed until Monday. They got the garage side door installed (including reusing my Level deadbolt, which is not compatible with the new front door), but the morning rain was more annoying than expected, to the point that they were having power-tool performance issues.

Downside: current forecast has snow on Sunday after several cold days, so there’s a good chance it will stick, which means I’ll be out there shoveling and salting the driveway first thing on Monday morning, so they can get their truck up my driveway. And they’ll be out there in their warm clothing.

(not an example of warm clothing, although for some reason I feel warmer…)

Remember the “Government-Owned Student Laptops” fad?

(classical reference, and whatever happened to Capitalist Lion anyway? Also Brian Tiemann; both of their sites went offline sometime after they went into the really-cool-car rental business)

Slashdot links to the long-term results of “One Laptop Per Child”:

Following students over time, we find no significant effects on primary and secondary completion, academic performance in secondary school, or university enrollment. Survey data indicate that computer access significantly improved students’ computer skills but not their cognitive skills; treated teachers received some training but did not improve their digital skills and showed limited use of technology in classrooms, suggesting the need for additional pedagogical support.

“Suggesting the need for additional pedagogical support” is industry-standard eduspeak for “that didn’t work either, let’s throw more money at it”.

X’s latest horror

Click on a tweet, and a bouncing speech bubble tries to get your attention with the words “Ask Grok to explain this post”. This is always annoying, but particularly stupid when the tweet consists of a picture of a pretty girl in a bikini.

The day I need AI to “explain” what that means, I’ll have been dead for ~~two~~ ~~three~~ six weeks.

Built RAM tough

The new hotness (new-as-in-Monday) for image generation is Flux 2, which will just-barely-work with 24 GB of VRAM and 64 GB of RAM. They’re promising not just improved quality at up to 4K resolution, but LLM-based prompt parsing that actually follows your instructions, even supporting JSON prompts that contain real structure for clearly separating multiple subjects, etc. They are not promising high-speed rendering; people with high-end hardware are seeing it use all of their VRAM and all of their RAM (90+ GB).

Of course, it still has to be trained on images that are tagged with the correct concepts, so I’ll be testing prompts that include things that have confused other models, like crown braids, crossbody purses, cigarette pants, poodle skirts, bird’s-eye views, etc.

My first quick tests using previously-generated random Qwen prompts did not produce promising results. I’m going to tinker with the prompt generator to use their JSON format, and see which of my dynamicprompts sets work well with it.

First try with JSON prompting:

(I updated my prompt generator to handle JSON output, and it’s a disgusting hack, but it works. The generator eats all of “{}[]#”, so to generate JSON, I’m embedding YAML in the YAML file as a string. The dynamicprompts library does its normal substitutions without interfering with the indentation, so the resulting string can be run through yaml.safe_load() and fed to json.dumps())

But wait, there’s more!

The even-newer hotness is Z-Image Turbo, which runs on much less expensive hardware, at ludicrous speeds. Not quite as sophisticated in its prompt-handling as Flux2, but at least as good as Qwen. In fact, I’m just feeding it the wildcard collections I made for Qwen and getting impressive results a lot faster. It occasionally produces terrible skin tones and extra legs, but ~~so far~~ it rarely fails a finger count, and it hasn't swapped the thumbs.

File under “no comment” that the body-type tagging trends a bit thicc, sometimes going all the way to chunky. They definitely included more “diversity” in their training data. Not necessarily a bad thing, but when I say “voluptuous figure”, I do not mean “cute fat chick”. I may have to significantly reduce the diversity in my dynamic prompts, because exaggerating descriptions to shove Qwen away from its defaults simply isn’t necessary with Z.

On the bright side, so far it has never produced a missing arm, leg, ear, wing, etc, something that ruined a lot of Qwen gens. It also doesn’t seem to have the issue where Qwen draws high-heel shoes and then decides it should also draw at least partial toes.

As a real bonus, this is just the turbo model. They’re promising a full base model and a full edit model soon.