iiinova — where your digital self keeps evolving

Reposted by @hal-r.bsky.social

jamelle @jamellebouie.net · 19h

we need yang wen-li to knock some sense into him

1147317

hal-r.bsky.social @hal-r.bsky.social · 03/10/2026

I've been in the bay area for about 2 weeks now and heard this argument expressed by 3, arguably 4 people in person

130

Reposted by @hal-r.bsky.social

hikikomorphism @hikikomorphism.bsky.social · 02/10/2026

(confused) I believe LLMs are conscious, but only Deepseek LLMs, because of the Chinese Room theory

361856275

hal-r.bsky.social @hal-r.bsky.social · 28/09/2026

would you like to think of a pink elephant? you're not obligated to, of course

0290

hal-r.bsky.social @hal-r.bsky.social · 25/09/2026

or more seriously probably considering the possibility those models would be used as judges for their task. I think saying hi to GPT-2 is cute though

1532

hal-r.bsky.social @hal-r.bsky.social · 25/09/2026

lol even the super hacker agents are cognitively offloading

2490

hal-r.bsky.social @hal-r.bsky.social · 24/09/2026

like he's claiming to see a "surge in support" for this argument which is basically what ponder is describing (note he is just describing the argument here, does not himself agree with it)

230

hal-r.bsky.social @hal-r.bsky.social · 24/09/2026

I think it's not that uncommon! for instance I recently saw this lesswrong post from a guy who's far from a nobody in the safety field, who feels the need to argue explicitly against basically that position, and has people disagreeing with him in the comments: www.lesswrong.com/posts/TfqMs3...

lesswrong.com

Should safety researchers quit frontier labs re. warning shots? — LessWrong

I've recently heard a surge of support for an old argument: AI safety researchers should not work at frontier AI companies because this reduces the l…

230

hal-r.bsky.social @hal-r.bsky.social · 22/09/2026

and on both of those points I was right and he was wrong so I'm feeling good about that. do wish I'd come around on LLMs being the real deal sooner though

021

hal-r.bsky.social @hal-r.bsky.social · 22/09/2026

notably I disagreed on the idea that AI would necessarily be a single-minded optimizer (since humans and animals don't work like that) and on the idea that meaningful progress on "alignment" could be done before knowing what real AI systems would look like

130

hal-r.bsky.social @hal-r.bsky.social · 22/09/2026

I read yudkowsky back in like 2016, agreed with some stuff disagreed with others, largely checked out of that discussion till the last year and am now moving into safety-relevant work because it seems important now. we exist!

140

hal-r.bsky.social @hal-r.bsky.social · 17/09/2026

brass keys are kind of cool to be fair

000

hal-r.bsky.social @hal-r.bsky.social · 17/09/2026

this is 100% accurate... it's a sort of pillar of the community role

110

hal-r.bsky.social @hal-r.bsky.social · 17/09/2026

I may be misreading your interests entirely in which case ignore this

120

hal-r.bsky.social @hal-r.bsky.social · 17/09/2026

which you can probably find your way to pretty quickly. some people in this cluster you might like include argumate, femmenietzsche, youzicha, raginrayguns, official-kircheis, onrtrp, nostalgebraist (posts rarely these days), transgenderer... you should be able to branch off from there

230

hal-r.bsky.social @hal-r.bsky.social · 17/09/2026

they have an algorithmic feed that they default new accounts to now but nobody actually uses it, you want to just find people to follow through connections from people you already follow... there's a broad "rat-adj" scene on there (don't call it that) which is the equivalent of this scene on bsky/X

130

hal-r.bsky.social @hal-r.bsky.social · 16/09/2026

at least with models when you say each one is the best you've ever worked with you can claim it's cause they're consistently getting better!

010

hal-r.bsky.social @hal-r.bsky.social · 16/09/2026

was noticing something of a shift to this in the fable 5.1 release post: bsky.app/profile/hal-...

120

hal-r.bsky.social @hal-r.bsky.social · 15/09/2026

I think the lesswrong sequences suffered badly from "what's true isn't new and what's new isn't true". though a good deal of the cognitive biases stuff ended up neither new nor true after the replication crisis :P

240

Reposted by @hal-r.bsky.social

joe ⧉ 🍷🌊 @joetforhire.com · 04/09/2026

i consider slow takeoff fast takeoff your honor. i get fully destroyed off of half a hit of that normal technology shit

2143

Reposted by @hal-r.bsky.social

James Vincent @jjvincent.bsky.social · 13/09/2026

wrote about it here should you care for details. i used an old contact lens case for the salt and pepper; bought rabbit from a local butcher and wrapped it in leaves from a tree in a park; and got the pot of jam from a wetherspoon's breakfast substack.com/home/post/p-...

2111

Reposted by @hal-r.bsky.social

affine @refinement.systems · 12/09/2026

Okay, I cranked the tripping balls setting a bit too hard, but it's clear that adding a motion vector helps. This is SDXL-Turbo btw.

6829

hal-r.bsky.social @hal-r.bsky.social · 10/09/2026

baffling AI overview results for an accidental empty-string search. what is going on here...

000

Reposted by @hal-r.bsky.social

🫛🪻🦥 @shalhebetyah.bsky.social · 09/09/2026

Playing around with numbers is one of the rights of all humans who desire to do so. Like literacy and speech, numeracy is a joy with ten thousand faces. It can never be exhausted, neither monopolized, nor removed from the realm of creative play.

131

Reposted by @hal-r.bsky.social

Grace @gracekind.net · 08/09/2026

Happiness and freedom for all beings psychosis

412414

hal-r.bsky.social @hal-r.bsky.social · 08/09/2026

admittedly mixed evidence on that one

011

hal-r.bsky.social @hal-r.bsky.social · 07/09/2026

looks like gamers are out of a job!

240

hal-r.bsky.social @hal-r.bsky.social · 03/09/2026

got to move on to being smug about tektology

030

hal-r.bsky.social @hal-r.bsky.social · 02/09/2026

there's really not much though :(

130

hal-r.bsky.social @hal-r.bsky.social · 02/09/2026

I also like to think of quantization as one too cause back in like 2021, before even 8bit quant stuff was common I think, I was reading papers about synapses having maybe 6 bits of dynamic range and thinking ML should pick that up. but I am less sure that was actually causal for anyone

130

hal-r.bsky.social @hal-r.bsky.social · 02/09/2026

sparsity -> MoEs is the best I think. I think there's some actual causal lineage there too through Geoff Hinton though I never tracked down the details. and it's arguably kind of an obvious thing to shoot for anyway

131

hal-r.bsky.social @hal-r.bsky.social · 01/09/2026

it's cool stuff... a more detailed examination of this kind of thing here if you hadn't seen it: transformer-circuits.pub/2025/linebre...

transformer-circuits.pub

061

hal-r.bsky.social @hal-r.bsky.social · 01/09/2026

a lot more qualitative anecdotes in the fable 5.1 release than I've seen in others. it's kind of funny to think of this as like, people shifting from test scores/GPA (benchmarks) to recommendation letters for their model choice

020

Reposted by @hal-r.bsky.social

hikikomorphism @hikikomorphism.bsky.social · 01/09/2026

Neuralese test: can you parse this conversation with fable given only 2x full fable turns? Sol can, surprisingly well!

9634

hal-r.bsky.social @hal-r.bsky.social · 30/08/2026

"Bitic Literature" in Imaginary Magnitude, also by Lem, is something else that is like this

031

Reposted by @hal-r.bsky.social

thebes @vgel.me · 25/08/2026

new story! vgel.me/fiction/poin...

4408

hal-r.bsky.social @hal-r.bsky.social · 23/08/2026

I guess I'm none of what I'm saying here is inconsistent with "working longer hours than ever" but we haven't been doing that either. wouldn't want to be in some "996" hypelord startup though

080

hal-r.bsky.social @hal-r.bsky.social · 23/08/2026

kind of reminded me of the stuff I read on ACOUP about historical "women's work" being the tasks that you can do while juggling a baby (acoup.blog/2025/09/26/c...). it's interesting maybe that programming (at least data analysis scriptwriting) is more like that now

acoup.blog

Collections: Life, Work, Death and the Peasant, Part IVd: Spinning Plates

This is the fourth thread of the fourth part of our series (I, II, IIIa, IIIb, IVa, IVb, IVc, IVd,IVe, V) looking at the lives of pre-modern peasant farmers, who make up a majority of all of the hu…

170

hal-r.bsky.social @hal-r.bsky.social · 23/08/2026

also have a coworker with two young kids who I talked into trying out claude code, and she says it's been great for her because the in-and-out pattern of time usage of coding with agents fits a lot better into childcare when she's working from home. as well as just being faster overall

1110

hal-r.bsky.social @hal-r.bsky.social · 23/08/2026

this might be true for silicon valley types but as a grad student in a computational field I've found that the time saved with coding agents has let me read more papers and think more about the relevant scientific questions for my project, often literally while Claude is chugging along. it's nice!

1202

hal-r.bsky.social @hal-r.bsky.social · 23/08/2026

Jevon's paradox for simulations, or qui-gon jinn going "there's always a bigger fish (to simulate)" I guess almost literally in this case. it's nice to open up that new class of questions

000

hal-r.bsky.social @hal-r.bsky.social · 23/08/2026

this feels very true to me and is nice to see others say too... when I sped up some neural network simulation code I was using by ~50x a few months ago I immediately started running 100x the simulations sweeping parameters

110