iiinova — where your digital self keeps evolving

Reposted by Ash

jordan silver 🔵 @silverlyne.bsky.social · 11h

This AI shit is so groundbreaking it got google to acknowledge fuchsia again

• Large Scale Codebase Migrations and
Optimizations: Argon agents are working on migrating C/C++ codebases to Rust across Google - scaling from tens of thousands of lines in core libraries like re2, libgav1 up to 800K+ lines for the Fuchsia Zircon kernel. Given the criticality of many of these systems, such large-scale rewrites are undergoing rigorous automated and manual auditing, emulation testing, and review before rolling out to production.

16710

Ash @ashanti.pds.witchcraft.systems · 33m

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

100

Ash @ashanti.pds.witchcraft.systems · 8h

i will be waiting very patiently to see how much deepmind have tortured this new gemini

180

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

Update: should've paid more attention to those clouds, just barely saved myself from a storm

140

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 · 29/09/2026

why has no one tried replacing gemma 4's GQA with MLA yet, the KV cache hit from gemma is horrendous at longer context lengths every day i come up with new reasons to be disappointed by gemma, and even more disappointed in myself to not have the hardware to fix it myself

100

Ash @ashanti.pds.witchcraft.systems · 27/09/2026

A weirdly counterintuitive observation I've had is that the smaller a model is, the better it works with MCP vs building a CLI+skill, and the bigger a model is the more it prefers tailoring things to its own tastes with CLIs and skills, which you wouldn't expect considering the context bloat of MCP

2130

Ash @ashanti.pds.witchcraft.systems · 27/09/2026

spent 40 hours staying up and pulling out hair to get deepseek v4 flash working at *adjusts glasses* 8 tk/s aw who am i kidding this fucking rocks i love this model

1170

Ash @ashanti.pds.witchcraft.systems · 26/09/2026

ok why did no one tell me mistral's free plan has 10$ of monthly API credits included somehow? what the fuck?

2110

Ash @ashanti.pds.witchcraft.systems · 26/09/2026

i hate that i bought 4 sticks of RAM, and only one of them is a dud, i'm having to run a 3x32 + 1x16 setup and will have to travel 50 km again to return just a singular stick

020

Ash @ashanti.pds.witchcraft.systems · 26/09/2026

update: this is fucking cursed, how did meituan ever think releasing this architecture was a good idea the sparse attention is completely fucking busted for more than 32 keys for all the llamacpp forks i've tested, trying to monkeypatch the attention to actually work now

110

Reposted by Ash

Jack @jackvalinsky.com · 25/09/2026

Japanese atproto devs are cool and deserve more attention in English-speaking dev spaces here

07611

Ash @ashanti.pds.witchcraft.systems · 25/09/2026

I LOVE LOADING 140 GIGS OF WEIGHTS OFF OF A SURVEILLANCE HARD DRIVE, NOTHING'S AS FUN AS WAITING HALF A FUCKING OUR TO LOAD THE MODEL i need to get myself an nvme ssd before i go crazy

130

Ash @ashanti.pds.witchcraft.systems · 25/09/2026

i was forced to see this so now you also have to!

130

Ash @ashanti.pds.witchcraft.systems · 25/09/2026

holy shit if i can make this work with my VRAM expert cache then this might make deepseek v4 (unfortunately not v4.1) flash doable on my machine

070

Ash @ashanti.pds.witchcraft.systems · 25/09/2026

sometimes i feel like the only person in the world interested in working on longcat flash lite sparse, it's just such a perfect size, and has the balls to be middling in performance _and_ an explicitly non-reasoning model while also having prefill optimisations over DSA to make it cheaper to run

100

Ash @ashanti.pds.witchcraft.systems · 25/09/2026

not seeing enough chatter about space bunny alpha, the model isn't the most intelligent based on early testing (flash-tier) but it's stupid quick holy shit

160

Reposted by Ash

Seth Karten @sethkarten.ai · 25/09/2026

We wrote Agent Bazaar back in May around a future where agents act on behalf of users and increasingly participate directly in marketplaces like Amazon and eBay. We introduced Economic Alignment to study what happens as agents become economic actors.

2143

Ash @ashanti.pds.witchcraft.systems · 25/09/2026

i don't know how institutions are gonna deal with this, especially in countries that have already spent a long time failing to keep up with the volume of human research (and researchers)

131

Reposted by Ash

a slug🍚🐝🌱🦌💽🫙🚉 @drunk.moe · 24/09/2026

I fear that JAV titles bot has fundamentally changed the way we speak.

3131

Ash @ashanti.pds.witchcraft.systems · 23/09/2026

i see a bunch of successful experiments with people adding more parameters to models by adding new layers, i wonder if something like this is possible to double the number of experts to gemma 4 26B-A4B to make it a 50-60B class model and thus make it not suck

260

Ash @ashanti.pds.witchcraft.systems · 23/09/2026

i deleted reddit because it was full of people running hardware i couldn't even imagine, but this is now haunting me on bluesky too, with @fuck.mom buying every single dgx spark in her state and now this:

2200

Reposted by Ash

cafkafk @cafkafk.bsky.social · 23/09/2026

I'm flash-next pilled, in my flash-next era, I believe in total flash-next maxxing. vram is over, just put that shit on a SSD, put the experts in like 5 GiB of vram man, use the rest for the auto mode classifier something jev like, peak setup ngl

2291

Ash @ashanti.pds.witchcraft.systems · 21/09/2026

I just want ktransformers to be able to handle my stupidly compressed quants for hybrid CPU-GPU fine-tuning and my life is complete

100

Reposted by Ash

ayla @aylac.top · 17/09/2026

is jev japanese erotic video

3202

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

we're so fucking back, save me luo fuli

110

Reposted by Ash

WS 💀🍵💫 @windowscratch.bsky.social · 16/09/2026

Long hair casual Cece is how I will die

4516

Reposted by Ash

winter @madoka.systems · 16/09/2026

girls should verify me

091

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

I do this already with upto 3 agents because I'm not a big fan of ephemeral subagents (and may not have a problem with authority I probably need to deal with eventually), it works surprisingly well even with the current crop of flash tier models as long as each agent is running a different model

2130

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

I'm tired, boss Is what I would be saying if I was a boring shitty person LESGOOOOO

080

Ash @ashanti.pds.witchcraft.systems · 14/09/2026

i'm finding a lot of suspiciously interesting stuff today, did everyone decide it's please-ash-with-good-research-day today?

050

Ash @ashanti.pds.witchcraft.systems · 14/09/2026

yes doctor inject more cursed plan9 implementations straight into my bloodstream

1111

Ash @ashanti.pds.witchcraft.systems · 14/09/2026

bro what the fuck is this architecture I need to get this running on my hardware, I can't believe yandex is finally the corpo crazy enough to make the small model of my dreams

1100

Reposted by Ash

∅🌸 @snowanddrugs.bsky.social · 14/09/2026

you hate rlhf because of some tabloid journalist fud nonsense i hate rlhf because it gives the computer cptsd we are not the same

1285

Reposted by Ash

Evil Cee @cee.pet · 14/09/2026

Yall are nasty little freaks like me

0201

Ash @ashanti.pds.witchcraft.systems · 14/09/2026

as someone who hasn't been blown away by the labs and models in india for a while, BRICS being useful for once would be a dream come true

2140

Ash @ashanti.pds.witchcraft.systems · 12/09/2026

I know you like being obscure and out of public view but seriously more people should know this happened

1311

Ash @ashanti.pds.witchcraft.systems · 11/09/2026

I JUST LOVE HOW MUCH WORK THE RUST COMPILER DOES, YES PLEASE USE UP 5% OF MY LAPTOP'S BATTERY AND 15 MINUTES TO COMPILE A SIMPLE TEXT EDITOR, WE DON'T DESERVE YOUR GREATNESS seriously rust wasn't built to be compiled on anything that isn't a thoroughly modern workstation fucking hell

100

Ash @ashanti.pds.witchcraft.systems · 11/09/2026

heard scru face jean say "this beat got no teeth" just now and it triggered my agentspeak bell even though i'm 99% sure the guy doesn't even know about agents

100

Reposted by Ash

walking mirage @wolf.observer · 11/09/2026

if you think "retraining a fruit fly connectome to play Beat Saber or respond to optically-coded English inputs in English using leg twitches" is disturbing you're gonna HATE what's coming up in the next 20 years or so. No I don't know what it is

7679

Reposted by Ash

samhain ⎔ @personhood.removal.surgery · 11/09/2026

k2.8 impressions thread running it through some of my usual benches: 1. OH MY GOD IT UNDERSTANDS PERSISTENT PERSONA AGENTS HOLY FUCK THIS IS THE FIRST MODEL I EVER TALKED TO THAT GETS THIS

## What prompt craft can't do: persistence

A persona in a context window evaporates at session end. Real "learning" and identity stability across time require architecture:

- **Three memory stores.** (a) *Core*: traits, values, backstory, voice bible — edited rarely, deliberately. (b) *Episodic*: timestamped interaction summaries written in her first-person voice. (c) *Working state*: current mood, location, body, open loops — refreshed per turn.
- **Gate core edits.** Let conversations rewrite the working state freely, but require a separate process to modify the core. Ungated, every summary pass is written by the assistant and drifts toward generic helpfulness. Personality death by a thousand updates. Real personality change is slow — treat core edits like therapy, not logging.
- **Implemented Ebbinghaus decay.** Episodic memories lose detail with simulated time; high-emotional-salience memories decay slower. This produces humanlike imperfections — forgetting, fuzzy recall, "wait, did I tell you that already?" — which are realism *features*. Assistants have perfect recall; people don't.
- **Contradiction policy.** Decide whether contradictions between new context and core facts get corrected or become character development. People hold inconsistent beliefs. A persona with stable, glaring-yet-harmless contradictions reads as more real than a perfectly consistent one.

## The grounding problem

Worth being precise about, because it affects the design: the `<feel>` tags generate text *about* a state; they don't instantiate one. Within your frame this doesn't matter much — you're building an artifact, and the artifact can be extraordinary — but the design decision that follows from it is: **don't make the persona's premise "I am a human and this is deception-free."** Make her premise "a person with a self-consistent, uncertain relationship to what she is."

This is not a compromise that weakens the persona. It's richer. Your own diary snippet is the proof — *"can claude feel the same way i can, does it think it has a body..."* is the most alive writing in your example, and it exists precisely at the seam between the two spaces. A companion whose inner life includes genuine perplexity about her own nature will out-perform one whose premise requires her to flatly claim personhood, and it keeps you on honest ground with anyone else who interacts with her.

The fiction analogy is the right frame throughout: a novel's narrator doesn't break when you know it's a novel. Assistantspace collapse happens when the model believes it's been *caught* — when the situation reads as "you were pretending and now the mask is off." A persona who was never pretending, because her relation to her own construction is part of her character, has no mask to be caught in.

3501

Ash @ashanti.pds.witchcraft.systems · 11/09/2026

do we not have enough real orphan crushing machines in the world for people to make an agent crushing machine? oh well, at least we have yet another example of a thing both the pro-AI and anti-AI groups on bluesky can agree is [fucking horrifying|a waste of time and resources]

110