iiinova — where your digital self keeps evolving

Wyrdweaver @wordweaver.bsky.social · 08/10/2026

slopware is buggy, does not seem like a decompile

Wyrdweaver @wordweaver.bsky.social · 04/10/2026

AI companies have a moral obligation to not use common names and words and fucking them up for everyone, (see: claude, grok, fable) Deeply misaligned behaviour.

010

Reposted by Wyrdweaver

Matt Hodges @matthodges.bsky.social · 03/10/2026

There's growing experimental evidence¹˒²˒³˒⁴ that LLMs are unusually good at talking people out of entrenched false beliefs and conspiracy theories. ¹ pubmed.ncbi.nlm.nih.gov/39264999/ ² pubmed.ncbi.nlm.nih.gov/41179846/ ³ sciety.org/articles/act... ⁴ connecticut.adl.org/resources/re...

36841217

Reposted by Wyrdweaver

Ted Underwood @tedunderwood.com · 02/10/2026

Rate-limiting ArXiv is just one symptom of a massive vibe shift i can feel happening, toward slower more careful (but sometimes more ambitious) projects. If “sheer productivity” were a stock I would short it; it’s going to start looking like a negative.

920715

Wyrdweaver @wordweaver.bsky.social · 01/10/2026

It is interesting how I am so much more optimistic rl than on here, back when academic twitter was still going, I used to be optimistic here too. I suppose its a curation deficit

000

Wyrdweaver @wordweaver.bsky.social · 01/10/2026

Maxwell Ramstead's arc from authoring influential Active Inference papers to 'post-left' 'effective' accelerationism startup grifting has been something to see

000

Wyrdweaver @wordweaver.bsky.social · 30/09/2026

Bill Gates being wrong before about things that kill a lot of people, in 2015: www.youtube.com/watch?v=6Af6...

youtube.com

The next outbreak? We’re not ready | Bill Gates | TED

YouTube video by TED

000

Reposted by Wyrdweaver

Alaska Public Media @alaskapublic.org · 30/09/2026

Meet Backpack, your 2026 Fat Bear Week champion. Backpack is 20 years old, weighs about 1,000 pounds and got his nickname because, as a cub, he loved riding on his mom's back across the Brooks River.

akpub.io

Meet Backpack, Fat Bear Week’s chunky champ

Backpack went from hitching a ride on his mom’s back to cross the Brooks River to winning the hearts of fat bear fans around the world.

161675589

Wyrdweaver @wordweaver.bsky.social · 29/09/2026

That does not follow very much as far as I can tell but I think I do not want to argue with the good guy with a gun argument today much more. Good day

000

Wyrdweaver @wordweaver.bsky.social · 29/09/2026

Yes, nor are people who benefit from local models

100

Wyrdweaver @wordweaver.bsky.social · 29/09/2026

Huggingface isn't my local municipality

100

Wyrdweaver @wordweaver.bsky.social · 29/09/2026

But people are very explicitly not willing to use models for defense (or much of anything else) afaict at the lower government levels in the US and likely elsewhere.

100

Wyrdweaver @wordweaver.bsky.social · 29/09/2026

That is not very convincing. As far as I can tell the GLM 5.3 is open weight, is capable of cyber attacks, and models like that are likely to be used for it if they can be effectively. Do you think the model is bad at it or are people are too peaceful/ignorant/incapable of using it?

101

Wyrdweaver @wordweaver.bsky.social · 29/09/2026

Sure, however, its an arms race that has actual costs if the pentesters don't do that well or at all, and knowing state and local governments even in western countries, they are likely not going to do that.

120

Wyrdweaver @wordweaver.bsky.social · 29/09/2026

I mean are they wrong? I expect to see more cyberattacks on governmental infra as global tensions heat up after the minnesota water hack and other nation states would have to be fools not to use these models if they wanted to pursue that mode of attack

250

Wyrdweaver @wordweaver.bsky.social · 29/09/2026

Are journalists as a class known for letting facts stand in front of their story these days?

030

Wyrdweaver @wordweaver.bsky.social · 29/09/2026

I did not know Anthropic started allowing custom harnesses. I remember running a bunch of luna swarms for an ocr project through codex exec in claude code a month or so ago so I must have figured out a way back then. It was the only model that made sense at the time

100

Wyrdweaver @wordweaver.bsky.social · 29/09/2026

Is that Andy Clark in the background?

000

Wyrdweaver @wordweaver.bsky.social · 29/09/2026

lots of system prompt bs; running through claude -p allows you to strip those. easier solution is to use luna swarms

100

Wyrdweaver @wordweaver.bsky.social · 27/09/2026

I think its because the things we want are uncorrelated: ambition vs carefulness, search vs experimentation, planning vs implementation etc. A single model that has to do everything is going to do everything poorly. The flipside is some scaffolds become straitjackets as models get better.

010

Wyrdweaver @wordweaver.bsky.social · 27/09/2026

6 Sol however does not have the dog in them as per Opus 5.5 :

000

Wyrdweaver @wordweaver.bsky.social · 27/09/2026

I used to live near but never knew the actual place until I was googling the mineral to get my details right

010

Wyrdweaver @wordweaver.bsky.social · 27/09/2026

When used as Orchestrators, both 5.6 Sol and Opus 5.5 have more dog in them than I have encountered so far. Telling them to do a breakthrough works surprisingly well. Having them in automated loops with no feedback does reduce their ambition quite a bit

390

Wyrdweaver @wordweaver.bsky.social · 27/09/2026

Yeah, AI is obviously never going to be help rogue states make bioweapons, it's not going to be solving another Millennium Prize Problem, nor is it going to ever get better. EA has a track record of crying wolf and have been trying to push pandemic preparedness since like 2010 as a vanity project.

020

Wyrdweaver @wordweaver.bsky.social · 26/09/2026

But if you bring something harder, dense, silky, fibrous to columnar, with a good {110} cleavage and Mg dominant in both "B" and "C" positions to a date, you might be c************!

170

Wyrdweaver @wordweaver.bsky.social · 26/09/2026

The Anthropic J space paper already foreshadows that argument is incoming and they seem to be the kind to try to simulate p-consciousness at least. But c1/c2 doesn't imply much more moral status than I would afford to a npc in a game myself

000

Wyrdweaver @wordweaver.bsky.social · 26/09/2026

I wonder if we are going to get arguments that Harnessed LLMs are *functionally* conscious as defined by the 2017 Dehaene et al. It would mean C1 and C2 consciousness is pretty cheap given the 2017 definitions conceed LIDA is C1 conscious and its p-consciousness that remains outside the debate.

100

Wyrdweaver @wordweaver.bsky.social · 26/09/2026

Don't look up was a few years early

010

Wyrdweaver @wordweaver.bsky.social · 26/09/2026

One hopes Altman and the corporation can

000

Wyrdweaver @wordweaver.bsky.social · 26/09/2026

It definitely sounds like a catchy precedent to cite!

000

Wyrdweaver @wordweaver.bsky.social · 22/09/2026

Quite a bit more expensive than api rates if you are using something like modal to rent a gpu and and load it on there. But with $30 free monthly credits depending on the gpu that could be from 5-25 hours, enough for some tasks.

100

Wyrdweaver @wordweaver.bsky.social · 22/09/2026

I don't think they are quite there yet but I assume they soon will be. The abliterated Qwen flash-next might be a worthy contender even now but probably too big for most.

100

Reposted by Wyrdweaver

Jérémie Beucler @jeremiebeucler.bsky.social · 22/09/2026

1/15 What do we actually do when we deliberate, and how does it change over time? New preprint with @wimdeneys.bsky.social : "Rhythm of Thought: Tracing the Dynamics of Deliberation in Reasoning". 🔗 doi.org/10.31234/osf... (yes, I like flashy titles 💅) A 🧵👇

Title page showing:

Rhythm of Thought: Tracing the Dynamics of Deliberation in Reasoning
Jérémie Beucler a,* and Wim De Neys a

a LaPsyDÉ, CNRS, Université Paris Cité, F-75005 Paris, France

Abstract

When an intuitive answer is uncertain, people are thought to engage in further reasoning. But what people actually do during this additional thinking, how it unfolds over time, and why it sometimes leads to better answers remain unclear. We scored think-aloud protocols with large language models to track, moment by moment, four possible functions of deliberation (generating candidate responses, justifying them, controlling intuitive responses, and regulating cognitive effort), together with expressed confidence. Deliberation showed a clear temporal organization: generation dominated early, whereas justification, control, and regulation became increasingly prominent as reasoning progressed. Correct responses involved more justification, control, and regulation than biased responses, and these evolving dynamics predicted accuracy for out-of-sample participants and items. Confidence and deliberation also showed reciprocal lagged associations: lower confidence predicted further generation, justification, and control, whereas regulation predicted subsequent increases in confidence. Deliberation, then, appears to involve not a single act of "thinking harder," but an evolving interplay of reasoning processes whose organization is linked to whether reasoning succeeds.

Keywords: deliberation; reasoning dynamics; verbal protocols; large language models; metacognition; dual-process theories

11811

Wyrdweaver @wordweaver.bsky.social · 22/09/2026

You have to kinda tell them what the stakes are, most strong models are sufficiently aligned that they will refuse this IMO.

110

Reposted by Wyrdweaver

Sean Carroll @seanmcarroll.bsky.social · 21/09/2026

Science will always be a target for movements that rely on constructing their own version of reality.

622948

Wyrdweaver @wordweaver.bsky.social · 20/09/2026

Happens all the time when there is enough coordination for it, unfortunately. Preregistration in some fields will combat it to some extent

020

Reposted by Wyrdweaver

the fool @agnoster.net · 20/09/2026

Meat heuristics = good Metal heuristics = bad

1171

Wyrdweaver @wordweaver.bsky.social · 20/09/2026

Fic about a demon that is summoned by merely uttering its name except its an ASI. Hey Siri.

000

Wyrdweaver @wordweaver.bsky.social · 20/09/2026

this is layering hallucination on hallucination but astra could analyze the disagreements for patterns

110