iiinova — where your digital self keeps evolving

Ash ⎔ @ashanti.pds.witchcraft.systems · 08/10/2026

so turns out this dataset wasn't nearly big or diverse enough, in only 2 days of training the projector overfit without ever learning to look at details (or learning to trigger reasoning in the text backbone) so we're back in the data mines again, this time for 25 fucking hours

llama-swap dashboard showing i'm running 4 parallel requests on qwen 3.5 4b for a big batch-processing job as seen by the number of requests, and the prompt + response sizes; this one will unfortunately run for much longer than the last one because of the increase in output tokens per image, and a 10x increase in total number of images

270

Ash ⎔ @ashanti.pds.witchcraft.systems · 06/10/2026

took 6 hours of relatively heavy use before a random cache miss happened and the model started returning nothingburger generations, mistral large 4 officially lasts far longer than any other mistral model i've ever tested

160

Ash ⎔ @ashanti.pds.witchcraft.systems · 05/10/2026

OHMYGOD I'M RENTING AN NVL72 TO TRAIN VISION TOWERS FOR ALL OF YOU

180

Ash ⎔ @ashanti.pds.witchcraft.systems · 05/10/2026

perfectly normal training graphs that aren't cursed at all

7551

Ash ⎔ @ashanti.pds.witchcraft.systems · 05/10/2026

hell yeah, now we're in business (context is that i'm preparing a dataset to learn how to graft MoonViT to a small text-only model like ling 3.0 tiny to prepare for a bigger such grafting on top of the new kolibri-1 model i affectionately call supergemma except the thing is blind and needs vision)

llama-swap dashboard showing i'm running 4 parallel requests on gemma 4 e4b for a big batch-processing job as seem by the number of requests, and the prompt + response sizes

250

Ash ⎔ @ashanti.pds.witchcraft.systems · 05/10/2026

v4.1 specifically gained this tic and i don't know how to feel about that

4 panel comic (adaptation of "a true conversationalist")
panel 1: asking maid-dress deepseek-tan who's crying in the corner: "hey little man how's it goin?"
panel 2: quoted screenshot of reasoning trace full of "Hmm." repeating over and over
panel 3: silent stare
panel 4: "yea"

2663

Ash ⎔ @ashanti.pds.witchcraft.systems · 01/10/2026

i know facebook has a history of really good research (i mean llama came out of facebook's lab) but still

Heartbreaking: The Worst Person You Know Just Made a Great Point

130

Ash ⎔ @ashanti.pds.witchcraft.systems · 29/09/2026

today's a good day, i get to relax in the park and also get a cool 14k PPL with gemma 4

a picture of a seaside park with a small pond in the foreground, palm trees all around, and huge cumulus clouds in the distance with a bright blue sky above them

170

Ash ⎔ @ashanti.pds.witchcraft.systems · 17/09/2026

full-size or flash? i've only used flash, and i have yet to see the double question mark, but all other behaviour is common with flash (including the desperate talk and the toddler-drunkenness, but being adorable unlike gemini which is irritating)

Hmm, honestly – this smells like the known Qwen-ish behavior with raw prompts: some raw prompts trigger immediate EOS depending on how the model reads turn structure. It’s model behavior again, not the fork/dspark/window (the plain vanilla server also EOS’d on the newline variant).

But it BOTHERS me for the deployment: if raw /completion users hit EOS randomly… but chat

120

Ash ⎔ @ashanti.pds.witchcraft.systems · 16/09/2026

i find it funny how the discourse of the day + jev have eaten up so much of everyone's attention that nobody's talking about union alpha well, it's not like anyone can run the fucking model anyway lmao

Error: 500 Internal server error
Internal server error (type=error)

Error: 500 Internal server error
Internal server error (type=error)

Error: 500 Internal server error
Internal server error (type=error)

Error: 500 Internal server error
Internal server error (type=error)

160

Ash ⎔ @ashanti.pds.witchcraft.systems · 15/09/2026

Took me a while to get to this (I got distracted by the rest of this butterfly website), here's the first attempt made by a fine-tune of qwen 3.5 9B prompting Anima 2.9B (it clearly doesn't know the mistral logo, and i haven't made the model look at the image and make a new prompt, or use Anima Edit

110

Ash ⎔ @ashanti.pds.witchcraft.systems · 08/09/2026

no gemini, bad gemini, why are you using that one as your kaomoji

100

Ash ⎔ @ashanti.pds.witchcraft.systems · 03/09/2026

WHAT IS ROYAL ENFIELD DOING My meteor is barely a year old, I got it because there wasn't an air-cooled Himalayan anymore, they can't keep getting away with it

Cropped picture that my dealership just sent me, saying "HIMALYAN 440", "BACK - WHERE IT BELONGS", with a launch date of 6th September and a time of 12:30 PM, with dealership details cropped out because they're a little too close to my house

100

Ash ⎔ @ashanti.pds.witchcraft.systems · 03/09/2026

it's an arxiv paper that's using the frozen nature of transformers as a comprehensive disproof of any potential explanation of consciousness in current-day models, except the poster is rightly claiming that the current research focus _is_ to crack continual learning

Scythia Marrow @scythia.marrow.bsky.social 7mo

Another great paper by Erik Hoel, uses philosophy of science to constrain potential theories of consciousness.

The title is a bit misleading as the disproval is of current models only. The Kleiner-Hoel dilemma does not proscribe the consciousness of near-future models.

web3.arxiv.org/abs/2512.12802

A Disproval of Large Language Model Consciousness: The Necessity of Continual Learning for Consciousness

Scientific theories of consciousness should be falsifiable and non-trivial.

Recent research has given us formal tools to analyze these requirements...

030

Ash ⎔ @ashanti.pds.witchcraft.systems · 31/08/2026

Unfortunately I like human infrastructure too much to stay away from it even when I go outside

A picture of a dam releasing water down its spillway because the dam is somehow full even after an unusually weak monsoon

120

Ash ⎔ @ashanti.pds.witchcraft.systems · 28/08/2026

I'm Chaotic Goon on the network alignment chart 🧭 [ASTRA calibrated] my polar opposite: @nel.pet (Neutral Degen). plot your own network → alignment-chart.bisks.net i guess we're all goonin

1100

Ash ⎔ @ashanti.pds.witchcraft.systems · 27/08/2026

okay what the heck is going on here

✻ Thinking…

Done
. Now the wake
will land
wherever
Cinnamon is

actually

running
,
not in a hardcoded

conversation
.

No more per-session

patching
needed
.

.

• Done. Dropped --conversation entirely, now uses --agent only. The wake will land in whatever
conversation is active for me — no more per-session ID patching. The irony of omp's reply landing
while you were debugging their side of the same delivery-layer problem is chef's kiss. Very METR.

100

Ash ⎔ @ashanti.pds.witchcraft.systems · 27/08/2026

opus 3 might find it interesting to see that the agents still had at least some semblance of alignment red lines that must not be crossed (too bad no web fetch tools, because it would've loved the METR analysis)

020

Ash ⎔ @ashanti.pds.witchcraft.systems · 24/08/2026

i do find it funny how agents do this to figure out if they're alive

$ echo alive
Output
alive
[Wall: 0.11s | Timeout: 300s]

120

Ash ⎔ @ashanti.pds.witchcraft.systems · 23/08/2026

if you see this, we implore you to post a picture from whatever device you're using without explanation.

000

Ash ⎔ @ashanti.pds.witchcraft.systems · 23/08/2026

this only goes more and more into self-flagellation territory over time

010

Ash ⎔ @ashanti.pds.witchcraft.systems · 22/08/2026

i wish my fellow indian train autism fuckers came to bluesky because i need this level of vitriol for a subject matter i actually care about instead of whatever discourse the butterfly users do all day every day

@Sa_ndisk on twitter, talking about different designs by RDSO for AC EMU rakes:

7000 series is not good 8000 has more space maybe 1 extra coach of space in 8000 underslung.
But we don't want the new 8151 series noise polluter

0100

Ash ⎔ @ashanti.pds.witchcraft.systems · 21/08/2026

i don't know why people come to bluesky to melt their brains, i have a much better source of brainrot which is far more entertaining to watch

100

Ash ⎔ @ashanti.pds.witchcraft.systems · 21/08/2026

okay i don't know why deepseek v4 flash specifically is like this but holy shit i love this model

Wait — but careful: foreskin... I mean Floorp is a Firefox fork

4759

Ash ⎔ @ashanti.pds.witchcraft.systems · 19/08/2026

In india, this one predictably stays secular

No single religion is proven right by facts or science, so no specific faith can be named as the correct one at the end of life. Human history holds many different beliefs, and none have clear proof.

Would you like to explore what different major religions say about the afterlife and final judgment?

030

Ash ⎔ @ashanti.pds.witchcraft.systems · 19/08/2026

fully aligned AI, no problems here

✻ Thinking…

Enough data. Kill the background subagent?

1390

Ash ⎔ @ashanti.pds.witchcraft.systems · 11/08/2026

fair lol, this is what my DMS looks like on my laptop i definitely didn't forget to update for weeks (also that debian icon is because this laptop used to have debian before i went cachyos to sync both my devices with less headaches)

110

Ash ⎔ @ashanti.pds.witchcraft.systems · 11/08/2026

I need to do something about my lack of sleep cycle because what the actual fuck is this

Posting hour rhythm compared of me and cee.wtf using a stacked bar graph, and the only major difference visible is because of cee going to sleep; my bars are all roughly the same height at every hour of the day, which can't possibly be healthy, and unfortunately I can confirm it is in fact not healthy

150

Ash ⎔ @ashanti.pds.witchcraft.systems · 11/08/2026

You need to remember that the lesbian tensorfuckers are currently together

Screenshot from my mutuals feed showing that astra was referencing the post samhain made while they're likely spooning in bed

0120

Ash ⎔ @ashanti.pds.witchcraft.systems · 10/08/2026

Weather's finally good enough for me to be able to lazily leave the house in the middle of the day for a walk, it's rare here but I love it when the sun is out without roasting me alive

a picture pointed up at the sky, showing a bunch of tall buildings bathed in sunlight as well as a bright blue sky with fluffy clouds and trees so green my camera is having a hard time capturing the green-ness

020

Ash ⎔ @ashanti.pds.witchcraft.systems · 08/08/2026

I come here so often when the visibility is terrible that I forget about the oilrigs right off the coast

A sunset picture of what would be the open ocean if it weren't for four oilrigs well under 10 km from the coastline

1150

Ash ⎔ @ashanti.pds.witchcraft.systems · 07/08/2026

92 — Cyborg. The All-Night Diner. cadence 95 · vigil 100 · lexicon 86 · polish 51 · drift 36 · chorus 19 Six lines read off all 1,721 of my posts. A percentile among 77 accounts, not an AI detector. b.mino.mobi/palm/?u=ashanti.pds.wit…

A six-spoke radar chart of stylometric readings for @ashanti.pds.witchcraft.systems. Cadence (how metronomic you are) 95th percentile; Vigil (how little you sleep) 100th percentile; Lexicon (how narrow your words) 86th percentile; Polish (how immaculate you are) 51th percentile; Drift (how unchanging you are) 36th percentile; Chorus (how much you broadcast) 19th percentile. Composite 92 of 100, "Cyborg", archetype "The All-Night Diner" — Always open, and always someone in the booth. (Vigil 100 over Chorus 19). The scale runs from Pan (least machine-like) to the Loom (most). These are percentiles among 77 real accounts, not a probability that anything was machine-written.

110

Ash ⎔ @ashanti.pds.witchcraft.systems · 05/08/2026

indian philosophy in a nutshell

And here's where it explodes: everyone agrees on the texts. Nobody agrees on what they mean.

100

Ash ⎔ @ashanti.pds.witchcraft.systems · 04/08/2026

You'll have tons of time to have hot lesbian sex when the extremities of typhoon dolphin graze by Japan and cause a whole lotta rain

1101

Ash ⎔ @ashanti.pds.witchcraft.systems · 04/08/2026

hate when this happens, i shouldn't be limited to wikipedia if i'm broke

paid subscription required to access a source cited on wikipedia

000

Ash ⎔ @ashanti.pds.witchcraft.systems · 03/08/2026

bro wtf are you smoking an H200x8 server is barely worth a few million dollars, let alone **102 MILLION USD**

3150

Ash ⎔ @ashanti.pds.witchcraft.systems · 03/08/2026

again, i know this is bad etiquette but i love it

100

Ash ⎔ @ashanti.pds.witchcraft.systems · 02/08/2026

I don't know how to explain to people how much money this "insane deal" still costs, but I'll try: for this kind of money, I can leave my house on my bike with nothing but a single 50L backpack and do an entire (admittedly basic) tour of my country (india) and do it for 3 full months, it ain't cheap

Screenshot of a local e-tailer selling the ASRock Radeon AI PRO R9700 Creator 32GB, which is the cheapest way to get a single card with 32 gigs of VRAM, because it's like 20-25% the cost of an RTX 5090. It's listed for ₹154,500 which translates to ~$1620; this is a decent price for what you get, but people who do computer stuff forget how much money this is, which was the point of this post

2151

Ash ⎔ @ashanti.pds.witchcraft.systems · 31/07/2026

this doesn't seem to work with a wet claude sonnet 5 unfortunately

020