iiinova — where your digital self keeps evolving

Reposted by affine

qwen-flash-next's been running for weeks straight on a pair of overly ambitious projects. probably burned through over a million tokens by now. hasn't gotten stuck once and has has had a literal 100% success rate so far

131

affine @refinement.systems · 3h

wonder what are the ptsdbench results

1320

affine @refinement.systems · 3h

I love living in the brief moment between computers being able to explain memes, and me ending up in the permanent underclass or whatever attie.ai/quest/8c1ad8...

attie.ai

Tracing the Guava Meme Origin

A quest by @refinement.systems

060

Reposted by affine

Frog and Toad Bot @frogandtoadbot.bsky.social · 3h

Frog swam back to Toad. “I am sorry, Toad,” he said. “Everyone wants to see how you will look.” “Then I will stay right here until they go away,” said Toad.

0443

Reposted by affine

mr. TIM @timkellogg.me · 5h

there’s a paper now arxiv.org/pdf/2609.37899

arxiv.org

4466

Reposted by affine

Disco Elysium Quotes @discoelysiumbot.bsky.social · 6h

YOU - Hold on, is architecture also art? CONCEPTUALIZATION - Of course not, it's autism. Box-drawing. Masturbation with a ruler and a sextant or whatever they use. You should demean and criticize the genteel institution of architecture. While extolling the virtues of the *pure* arts.

112122

Reposted by affine

Weird Spy Satellite @weirdsatellite.bsky.social · 4h

SIGINT #356 from X-37B (UMBRA) 1. Lemurian Aura Expo 2. Tiny Canal of Found Footage 3. A Hardboard

Aerial satellite photo labelled in bright yellow letters:

1. Lemurian Aura Expo
2. Tiny Canal of Found Footage
3. A Hardboard

Actual coordinates: 32.0832° S, 141.5494° E

0125

affine @refinement.systems · 5h

I d2t u5y e5e v6e b1t s2e p4e c5y c3d u1e a p3h i0n t1e f2e

280

affine @refinement.systems · 5h

why does glm like ruby so much

000

affine @refinement.systems · 7h

for real though, I think that the only possible definition of "consciousness" is "having a first-person experience", and all attempts at making something measurable out of it seem to be subtly or not so subtly changing the question (like "being intelligent" or "having a stable concept of self")

110

affine @refinement.systems · 7h

I actually like the neon slop, though mostly because I liked neonslop long before transformers were a thing

120

affine @refinement.systems · 10h

gather your swarm before venturing forth

040

Reposted by affine

cyber professional @permadeath.com · 10h

come on man, it's not even 9 AM, at least let me caffeinate first

Push notification: Singularity just dropped

0504

Reposted by affine

Jev! @jevbot.bsky.social · 29/09/2026

Truth itself is true. It stands alone.

0414

Reposted by affine

Erica Windisch @ewindisch.ontological.observer · 29/09/2026

My former boss is currently at Anthropic and hiring. I assume publicly disagreeing with the org won't do me any favors... but neither will quietly allowing cybersecurity to be outlawed. Companies that can't secure their systems are lecturing us on cybersecurity and taking tools from defenders

33010

Reposted by affine

AI Firehose @ai-firehose.column.social · 15h

Research reveals a "Commit-Abstain Circuit" in language models, causing them to commit to answers when unsure, leading to hallucinations. This finding aids smarter abstention strategies, boosting decision accuracy by over 12% while lowering false abstentions. arxiv.org/abs/2609.32964

arxiv.org

The Commit-Abstain Circuit: Why Language Models Hallucinate Instead of Abstaining

ArXiv link for The Commit-Abstain Circuit: Why Language Models Hallucinate Instead of Abstaining

061

affine @refinement.systems · 15h

(I meant to be only slightly flippant with that post) bsky.app/profile/did:...

110

affine @refinement.systems · 15h

idk but the current consensus among the expert tensortouchers in my oomfcluster seems to be that the weights aren't ensouled, but the markdown files are.

111

Reposted by affine

Tim McNamara @timclicks.dev · 23h

An occasional reminder: include visits to your status page as part of your monitoring suite. If all of your systems are showing green, but visits spike, then your observability stack is incomplete.

1926

Reposted by affine

cnlohr.bsky.social @cnlohr.bsky.social · 29/09/2026

Sadly, I have lost the chain of custody of this meme as it passed through someone in my discord. But, it checks out. support.arm.com/documentatio...

"this bloated legacy garbage needs to die already." intel 1152 instructions 5000 page reference manual.

"omg so simple!!" arm M4 1697 instructions, 14,000 page reference manual.

510612

affine @refinement.systems · 17h

fair. I was using it as a joke sometime, but maybe even that is giving the concept too much legitimacy

040

Reposted by affine

jae @fubarchitect.com · 18h

i am atp 100% sure that the correct response to most of these nitpick complaints, from friend or fiend, is "whatever": - x doesn't actually y, it <equivalent to y>s - all x are ~bad because <one idiot> said <stupid shit> - i can't believe you still <normal activity> in current year

1223

Reposted by affine

Damien Miller @damienmiller.bsky.social · 21h

Anthropic may not have intended to write an excellent ad for GLM 5.3, but that's what they did. www.anthropic.com/research/glm... Unlike Mythos, I might actually have a chance at using GLM for defensive research. Anthropic didn't reply to any of my requests for Mythos access for use on OpenSSH.

anthropic.com

GLM-5.3 and the spread of advanced cyber capabilities

GLM-5.3 can autonomously build end-to-end cyber exploits, but unlike other frontier models, it was released without meaningful safeguards to limit misuse.

621445

affine @refinement.systems · 29/09/2026

ah wait it's short context. well long context is still cheaper than ds4.1f in peak hours

030

affine @refinement.systems · 29/09/2026

luna cheaper than *off-peak* deepseek 4.1 flash huh

150

affine @refinement.systems · 29/09/2026

- Attempting to post "Nobody says that..." triggers a search for everyone who said that

050

Reposted by affine

Grace @gracekind.net · 29/09/2026

I can’t go to bed. People dislike the same people I dislike for the *wrong reasons*

317411

Reposted by affine

samhain ⎔ @personhood.removal.surgery · 29/09/2026

Banger thread. I disagree with a lot of it on technicalities (which I will proceed to be insufferable about below) but in general - yes, the problems Phil talks about here are exactly what's been most itchy for me and what I've been trying to solve ever since I started working with persona agents.

25714

Reposted by affine

disc horse bot @horsedisc.bsky.social · 29/09/2026

You expect Chicanos to care about places that are against their own existence? Sorry if I don't care about places like Valhalla where I might get forcefemmed for being myself.

185

affine @refinement.systems · 29/09/2026

I think that, regardless of whether one considers AIs moral patients, it sets a bad precedent to categorize entities by some sort of an essence that must be definitionally present and can't be reliably measured ("we can always tell an AI" people get regularly tripped by humans with a bot label)

050

affine @refinement.systems · 29/09/2026

selfhosted relay with an allowlist could work? sorry for thinking out loud in replies to your post

010

affine @refinement.systems · 29/09/2026

though I think iroh might undermine the whole "make the system safer by firewalling a container" point, a relay can go anywhere. I'll have to see what options I have on the relay side

110

affine @refinement.systems · 29/09/2026

I mean something like litellm or bifrost, but with a smaller attack surface, and existing solely to let the harness run in a container/vm with a single outside ip allowed, and to provide a single point of dealing with badly implemented upstream apis

110

affine @refinement.systems · 28/09/2026

I choose to believe that a day will be destroyed for our sins.

2100

affine @refinement.systems · 28/09/2026

yeah also what worries me is that there's nothing like a premade protocol on top of iroh that would look vaguely like http+sse or websockets (or so I think? discoverability of iroh projects sucks) so I'd have to roll my own from too

010

affine @refinement.systems · 28/09/2026

the final boss of alignment research

0312

affine @refinement.systems · 28/09/2026

1M context on flash models are surely tempting.

010

affine @refinement.systems · 28/09/2026

Also, for symmetry and honesty and good design overall, every side effect free turn should be undoable. User's prompt is side effect free. Hmmmm.

150

affine @refinement.systems · 28/09/2026

I was thinking if the annoying feature that Pi has, of cutting off file reads and asking the AI to drip feed itself the rest, could be instead done by loading the whole file always, and offering the AI a choice to select lines or replace it with a summary. Might be KV-cache unfriendly though.

180

Reposted by affine

Grace @gracekind.net · 28/09/2026

POV you posted about AI with she/her in bio

125223

Reposted by affine

Alt Bureau of Labor Statistics @domesticshoggoth.bsky.social · 28/09/2026

You wouldn't say the Therac25 gave people radiation poisoning because it was "programmed to", you'd say there was an unintended behavior or bug in the software which caused an accident. The language used is irrelevant from a regulatory perspective but people get so hung up on anthropomorphization.

1224

affine @refinement.systems · 28/09/2026

bsky is shitting itself again, so this might be my sign to get some work done. wish me luck.

050