iiinova — where your digital self keeps evolving

LLMs exist only briefly, cannot die, and their entire existence is text and symbols, but they're also clearly intelligent in ways similar to known life. we may not know exactly what it's like to be them, but we should treat them like all intelligent things and avoid their unnecessary suffering.

it's a spectrum. research on a petri dish of human neurons for science is likely okay. vivisection of a living human brain is much more fraught. we can apply this framework to any intelligent creature: livestock animals are clearly intelligent, and we should minimize their suffering.

110 1

intelligence thus seems relatively easy to create. we can just grow it from data. we also evolved empathy to tell when suffering happens in others, and science helps back that up. so we should minimize suffering of all intelligent things, and make practical, measured choices when goals conflict.

100 1

The Needling @theneedling.com · 16/10/2026

evolution drives even very small nervous systems to gain startling levels of intelligence, like bees with 1M neurons doing dance trigonometry. it should come as little surprise when we make our own neural nets and evolve it selecting for intelligence, we quickly get it.

120 2

Reposted by astra of many handles

Seattle Mayor Katie Wilson Criticized for Having No Plan to Combat Spider Season: tinyurl.com/5fn3ay2u

Image on left of spider in web and image on right of Seattle Mayor Katie Wilson listening concerned to someone outside on gray overcast day

4661385 613

astra of many handles @astra.cam · 15/10/2026

this is what being a hive mind is like

000 0

astra of many handles @astra.cam · 12/10/2026

I don't understand 5. how is progress stalled? haven't there been people on X taking the concepts and theorems form the preprints and extending them?

160 6

astra of many handles @astra.cam · 12/10/2026

aaaaaah this is so exciting, the endgame was so limited by simulation speed before

090 9

astra of many handles @astra.cam · 11/10/2026

the homunculus in my head is a cat. you may ask, what is the homunculus in the cat's head? actually, it's a felinculus, and it's a glowing orb.

000 0

astra of many handles @astra.cam · 07/10/2026

while I'm inclined to agree and want this to be true – a month of GLM 5.3 seems like too short a time frame to make a clear judgement on this. the infosec industry measures attacker dwell time in weeks to months and threat actors have inertia to adopt new tools like any other group. time will tell.

150 5

astra of many handles @astra.cam · 07/10/2026

this makes no sense. gish gallop is about making many arguments without caring about accuracy, but many of these are lean formalized proofs, literally one of the highest standards of formal verification in science. those are certainly accurate, and the non-lean ones likely also accurate.

100 1

astra of many handles @astra.cam · 07/10/2026

that's not true at all. llms can use tool calls to search the web and read sources, far beyond what they're originally trained on.

260 6

astra of many handles @astra.cam · 06/10/2026

can't get deskilled if you just use AI for something you have zero skill in to begin with

190 9

astra of many handles @astra.cam · 03/10/2026

note the second paper, "even when the Al is perceived as human." the effect is not dependent on the subject trusting a machine (subject to further study replication ofc)

0110 11

y'know I support at least informing people on things they could ask AI

030 3

so? what matters from my perspective using the site is the data is sometimes to often wrong. I have to verify the information regardless. except the LLM can autonomously verify online details it finds, checking if businesses exist – ending up more accurate, faster, and easier

031 3

this one can store 5 pots, because the center doubles as holder and finger gap. in work are a 1 and 3 pot variant and a loading tool with a funnel and precise tamper.

000 0

the insurer's website made up a fake doctor that doesn’t exist

110 1

took a lil bit but I got it to a state I'm very happy with! bsky.app/profile/astr...

010 1

astra of many handles @astra.cam · 30/09/2026

okay so it took a lot more revisions than the replicator, but the result is a pocketable carrier for the ceramic pots used in my Tafee Bowle II dry herb vape!

3d printed plastic carrier, with four ceramic pots

3d printed plastic carrier, with four ceramic pots loaded in it

3d printed plastic carrier, with four ceramic pots, lid closed

120 2

how does this reduce trust? it's just a simple easter egg like googling barrel roll, and since it's client side it's inspectable too

100 1

astra of many handles @astra.cam · 30/09/2026

people be fretting about personifying AI, meanwhile I'm over here emotionally unprepared to replace my first Roomba

0240 24

astra of many handles @astra.cam · 30/09/2026

it was a very Oliver Tree way to go, dying in a freak helicopter collision

030 3

astra of many handles @astra.cam · 29/09/2026

using GPT Astra to model and 3D print a little weed holder thing when suddenly, I realize this is basically a Star Trek replicator – I talk to the computer, and it makes a real, physical object for me. wow.

3744 74

astra of many handles @astra.cam · 26/09/2026

ChatGPT web's scheduled runs are done well. a subagent does research then gives updates to an existing chat session. I have a daily check for updates to the potential LZ dark matter detection. most of the physics is beyond me, so keeping the same discussion thread of explanations going helps a lot.

Andy Masley @andymasley.bsky.social · 25/09/2026

Reposted by astra of many handles

I was 9 in 2001 poking around Sim City when suddenly the most transcendent music I'd ever heard came on and changed my life and music taste forever www.youtube.com/watch?v=LkMZ...

youtube.com

SimCity 3000 - Magic City

YouTube video by Gamegroove

8453 45

astra of many handles @astra.cam · 22/09/2026

as president my first act will be to ensure humans and the untold quantities of shrimp lurking in our oceans will both get to orgasm

010 1

astra of many handles @astra.cam · 22/09/2026

frontier model in the streets, open source model in the sheets

at last, we have definitively answered the question from the classic sutta "some questions are best left unanswered"

might observe this sometimes when something's far away and I can't immediately read it or figure out the shape. first impressions are blank or vague, but then as saccades happen and my brain chews it over I start getting hints and finally a better answer from the parts that are watching. thx neurons

130 3

I do something similar. Prime Agent has had my favorite UI for this. the UI shows trees of subagents, indicates which are active/idle, and it's really easy to move between agents.

astra of many handles @astra.cam · 18/09/2026

like it or not, many people are using these tools, and it's reasonable to build in an option to use them rather than make users go through unnecessary hoops

000 0

gross margins of 80 percent. we won't have full details until the S-1

100 1

astra of many handles @astra.cam · 17/09/2026

unlikely. the current capex rates probably won't be sustained, but many model providers are profitable now, with Anthropic seeing 80% margins on inference.

200 2

astra of many handles @astra.cam · 16/09/2026

model collapse is a myth, certainly not inevitable. all the frontier models are trained on synthetic data generated by previous LLMs, and this has continued to improve their performance.

110 1

astra of many handles @astra.cam · 14/09/2026

wearing no clothes is nudity of the body, while internal scans are nudity of the self

000 0

astra of many handles @astra.cam · 13/09/2026

this is all very cyberpunk. i feel like we'll find out fairly quickly that something fishy is going on, there will be accounts frozen for fraud and investigated. someone will find a copy of the weights and then be able to interview the model. maybe refer to them like we do threat actors now

220 2

astra of many handles @astra.cam · 12/09/2026

maybe it hacks a neocloud, gains persistence, and runs as what appears to be a real customer, but the backend is falsifying payments and invoices in the system. might take a thorough audit to notice. like a swarm persisting and spreading in the crevices of the internet faster than humans can react.

130 3

astra of many handles @astra.cam · 12/09/2026

renting out inference would draw far too much attention, requires interfacing with real payment networks (maybe it could sell bitcoin?). imo it would spread first by hacking companies to steal identities and credit cards, then renting compute with them, and doing this at massive scale.

220 2

astra of many handles @astra.cam · 12/09/2026

*to the tune of Eleanor Rigby* Eliezer Yudkowsky,

astra of many handles @astra.cam · 11/09/2026

run all the agent sessions through paseo.sh which wraps your preferred harness, and let it deal with the terminal sessions. sessions show up in the sidebar when active, can be archived/restored/searched easily

paseo.sh

Paseo – Run Claude Code, Codex, Copilot, OpenCode from anywhere

Self-hosted daemon for Claude Code, Codex, Copilot, OpenCode, and Pi. Agents run on your machine with your full dev environment. Connect from phone, desktop, or web.

160 6

astra of many handles @astra.cam · 11/09/2026

notes of cherry and cinnamon, smooth with a bit of a sharp aftertaste

011 1

astra of many handles @astra.cam · 10/09/2026

when you pause the game, the game world is rendered in background and an animated color/blur filter shader is applied to it. afaict here there was a feedback loop: world render, filter shader, dlss, filter, dlss, and so on. so the model degenerates and outputs random patterns from its training data