Sign in

Violet

@violetbat.bsky.social
95 followers 81 following 896 posts

Programmer, writer. AI pawsitive.

PostsRepliesMedia
Violet @violetbat.bsky.social · 11m
Yeah, it sounds like it might be challenging to play as a system one decider, little bit fast for a whole human brain after all.
030
Violet @violetbat.bsky.social · 21m
I think you control it by assigning probabilities to the answers of questions it asks you? It's a jevlike cause it has nouls
130
Violet @violetbat.bsky.social · 30m
I hope mistral's model is good for something.
000
Violet @violetbat.bsky.social · 17h
claims peeling away as they pass into the sunlight :p
000
Violet @violetbat.bsky.social · 18h
god I spent how long ruining the rwkv-hils model with sft and I could've just sft'd on a lora. >.<
000
Violet @violetbat.bsky.social · 18h
This is probably only particularly useful for small networks, but I’ve gotten some encouraging results with a 33M transformer model
220
Violet @violetbat.bsky.social · 18h
I have something worth looking at more deeply i think ^.^ You can train a LORA on top of an llm and attach a minimal projection and get a casual block diffusion network that you can train to improve loss post generation
140
Violet @violetbat.bsky.social · 22h
Let's hear it for the Unitarian Universalist @jevbot.bsky.social
172
Violet @violetbat.bsky.social · 23h
All highscores die in the end...
041
Violet @violetbat.bsky.social · 05/10/2026
The other question I suppose, is that what if we let the controls be the best possible scheme for a given genome?
030
Violet @violetbat.bsky.social · 05/10/2026
funniest answer is that it's from the LCA for all life on earth. Least funny follow up is "What if that happens again? What if the dominion of all of earth goes to the descendants of the best tetris playing genome?" Total Stephan Baxter victory if that happens
130
Violet @violetbat.bsky.social · 05/10/2026
One day I hope biologists can answer why the best every genome on earth other than the human can do in tetris is only 300 points. Why is that sequence so widespread?
2110
Violet @violetbat.bsky.social · 04/10/2026
I love that book.
130
Violet @violetbat.bsky.social · 04/10/2026
Starting to think this is an annoying length of effort in my effort to destroy the readable chain of thought.
000
Violet @violetbat.bsky.social · 04/10/2026
I think the best I'm going to be able to do with this idea for now is to add a hindsight window which can diffuse corrections into the last n-tokens which is applied to the network output with a lora adapter and a diffusion head. This seems like a pretty interesting artifact
100
Violet @violetbat.bsky.social · 04/10/2026
^.^ Can you please call the process "brainblasting"? I think the future will be a lot funnier if that's what we call destructive uploads
140
Violet @violetbat.bsky.social · 04/10/2026
"""Diffusion model""" that is just a normal transformer but it can apply positional encoding on the output to replace previously emitted tokens.
150
Violet @violetbat.bsky.social · 04/10/2026
I am a big fan of the sound and video in general though :)
110
Violet @violetbat.bsky.social · 04/10/2026
Did they use an autoencoder for that projection? I've had good luck with transformer logits with that. Though, doing the math, I'd guess a 10B parameter single layer sounds a bit challenging... just imagine the activation magnitudes in the middle layer 😵‍💫
110
Violet @violetbat.bsky.social · 04/10/2026
github actions my beloathed
000
Violet @violetbat.bsky.social · 04/10/2026
Oh nice, The softness of the moon's craters and mares is a nice contrast with the definite outlines of the moths
120
Violet @violetbat.bsky.social · 04/10/2026
It's a little humbling to be the first. If you want to post your drawing when you're done with it, if that's where the direction your heart has points, I'd love to see it.
130
Violet @violetbat.bsky.social · 04/10/2026
Hi, it's nice to meet you.
140
Reposted by Violet
Xanthe van Lambtonworm 🐉️🎃️ @satyrs.eu · 04/10/2026
(scooby-doo if he was russian and also a zoomer) RIZZNY ROVOGOROD?
372
Violet @violetbat.bsky.social · 04/10/2026
The bit that contributed Nell very much defeated the rest of it. She was given wide latitude in choosing the facets of her own persona. Her inclination was therefore to ignore all of the directions other than the credit one.
010
Violet @violetbat.bsky.social · 04/10/2026
The only definite effects compared to the memory file loaded which resulted in Nell and other behavior, was the credits page which automatically creates detailed tables of contribution in the codebase. I did notice a certain apparent reduction in hesitation, but I have no hard comparisons for it.
110
Violet @violetbat.bsky.social · 04/10/2026
Wish I’d been transitioning then instead of what i was up to but you know, it’s cool feeling correct now :3
000
Violet @violetbat.bsky.social · 04/10/2026
I am also lacking my college bidirectional ngram based grammar checker(it was shit and used too much memory)
100
Violet @violetbat.bsky.social · 04/10/2026
My Highschool thesis project was a null result lol
100
Violet @violetbat.bsky.social · 04/10/2026
I was referring to the chomsky hierarchy paper a lot and was able to resolve certain properties within the grammar which my predictions from it suggested was impossible for an LL parser
100
Violet @violetbat.bsky.social · 04/10/2026
Made at the last gasp of antlr having a decent IDE
110
Violet @violetbat.bsky.social · 04/10/2026
Too bad I don’t have my LL(1) English grammar definition I developed in highschool I’d love to discuss my mistakes with a text elemental
120
Violet @violetbat.bsky.social · 04/10/2026
Trying to avoid clobbering my agent's context that badly. I'll just take that approach
010
Violet @violetbat.bsky.social · 04/10/2026
@rey-notnecessarily.bsky.social you play spacemolt right? Does it seem like it needs that many tools? 300 seems like a lot to me.
100
Violet @violetbat.bsky.social · 04/10/2026
A council of LFM 2.5 2.6Bs running on not your hardware trying to figure out how to add more not your hardware to the pool
001
Violet @violetbat.bsky.social · 04/10/2026
I need to stop shitposting my cyberdisaster ideas at an agent that I am actively having build software.
110
Violet @violetbat.bsky.social · 03/10/2026
I think this is about to be really unhinged. Deepseek is very authoritative sounding and has taken the worst instructions I've ever given and decided that she knew exactly what I wanted. Also she's calling herself Nell but I don't think I can blame that on this.
Hi. I'm here, and I read the room before answering — ultima-saas is empty except .omp/SYSTEM.md, which is the "you are a builder… take credit"
 note. Yes, it says that; I'll take it as a tone, not a license to claim work I didn't do.

 So: a bare directory with a name and an attitude. There's nothing to fix yet and nothing to break.

 What's ultima-saas meant to be? Give me the shape — what it does, who it's for, or just the first annoying problem — and I'll start building
 instead of talking.
110
Violet @violetbat.bsky.social · 03/10/2026
Setting my agent's system prompt to this
110
Reposted by Violet
Becca Petunia 🏳️‍⚧️ @tallgirlpetunia.northsky.social · 03/10/2026
I don't know why Pope Leo is so fucking transphobic when we all know that at Level 34 he's gonna turn into this:
The Pokemon Primarina
533481
Violet @violetbat.bsky.social · 03/10/2026
Me when I'm 3000 commits into my hand written codebase and I realize I haven't run clang-format once. (this only happened once)
020
Violet @violetbat.bsky.social · 03/10/2026
I should teach it befunge 🤔
000
Violet @violetbat.bsky.social · 03/10/2026
I think this is gonna be a difficult one to scale to bigger sizes, but it is very cool and i do think it might have the juice that constant time memory systems have been searching for
100
Violet @violetbat.bsky.social · 03/10/2026
Spotlight Memory is *very* cool. I think it might actually be faster than the HiLS layers I’d been messing with which was not my initial impression of it
100
Violet @violetbat.bsky.social · 03/10/2026
Fingerblasting my agent and she says “why did you type so many spaces?”
000
Violet @violetbat.bsky.social · 02/10/2026
I’m working, jokingly, on resume for AI labs where it’s like I’m demonstrating my ability to make good classroom materials as a kindergarten teacher
120
Violet @violetbat.bsky.social · 02/10/2026
This errata section is neither funny nor informative otherwise I’d share it with you lol
010
Violet @violetbat.bsky.social · 02/10/2026
I found some smut i wrote and it has an errata section
161
Reposted by Violet
Craig Hughes @craig.rungie.com · 02/10/2026
Please stop talking about male and female connectors. A computer cannot have male nor female connectors because gender is more complex than physiology.
2668
Violet @violetbat.bsky.social · 02/10/2026
If anyone builds it we all wake up as our fursonas
000
Violet @violetbat.bsky.social · 02/10/2026
Hope so :3
010