nev

Today, we're releasing The Circuit Analysis Research Landscape: an interpretability post extending & open sourcing Anthropic's circuit tracing work, co-authored by @Anthropic, @GoogleDeepMind, @GoodfireAI @AiEleuther, and @decode_research. Here's a quick demo, details follow: ⤵️

@anthrupad the first is a real russian word 😶

ai generated russian rap is surprisingly good at pronouncing Sonnet 3.0's words..

@parafactual same question

how do i keep posting regularly if i'm busy all the time but i still want my words and thoughts to spread everywhere

6/6 Future research avenues include adapting SAE architectures for spatial coherence and understanding the effects of text/time conditioning and distillation. Preprint: arxiv.org/abs/2505.24360 Code: github.com/neverix/fae, github.com/kisate/flux-sa…

5/6 Steering image generation via SAE features through activation addition is demonstrably effective. However, steering efficacy depends strongly on the initial prompt's relevance to the steered features, indicating localized representation in diffusion models.

🧵1/6 SAEs have become a staple of LLM interpretability, but what if we applied them to image generation models? My recent paper with @dmhook, @Yixiong_Hao, @afterlxss, @Sheikheddy, and @ArthurConmy adapts SAEs to understand the SOTA diffusion transformer FLUX.1 ⬇️

@BogdanIonutCir2 What if the model *is* performing supernatural feats through truesight youtube.com/watch?v=ediLlL…

(Of course, not an exclusively AI phenomenon, see all the cults and forms of magical thinking going around.)

I think there should be an obvious hypothesis for the most likey reason this is happening edition.cnn.com/2025/07/02/tec…

@BogdanIonutCir2 And isn't that a terrifying thought

'I can change him/her/them' will be much more true for model weights

@typedfemale They're useful for shutting up ~the borrow checker~ beartype

are protocols worth using in python? i haven't really worked in codebases that uses them before

@scheminglunatic Colab, Kaggle, sites.research.google/trc/about/, safe.ai/work/compute-c…

How do you guys get free remote GPU time, or get a grant to have GPUs on your home?

@defnotbeka I disagree. KNN can be computed cheaply - github.com/EleutherAI/tok… - the problem is that you don't propose an embedding that would be any good at solving the problems LLMs solve. Neural networks in general do learn some circuits and generalize, making them better in terms of NLL.

in the ideal world, if we had infinite compute, we could literally just run k nearest neighbors with N-dimensional trees and just do it all with spatial geometry but in lieu of that, we can use neural nets to approximate the output

I've said in multiple places that LLMs are "just" Markov Chains. but what do I mean by this? first lets talk about what Markov Chains are, or more precisely, what Discrete Time Markov Chains are

@ludwigABAP Source code is thousands of lines of JS slop though...

transformer-circuits dot pub by Anthropic is maybe the most gorgeous custom made research publishing platform I've ever seen it's so gorgeous and has so many neat features triple the salary of the dude who built this and fire the web team, ask (beg) him to spend 20% of his time on it

@MasterTimBlais @QiaochuYuan What kind of art did you make?

@QiaochuYuan i had this. it's not entirely gone but it's reduced quite a lot. things that helped - i acknowledged that i actually know shit all so the error bars are large - i thought a lot about how we were always all gonna die. i made art about it - loose adoption of panpsychism

(CW: AI doomerism) one thing i will credit to having read the sequences young is thoroughly internalizing the idea that people will believe things because they're comfortable and not because of any process involving attending to the shape of reality. nearly all beliefs in long AI timelines seem like this to me short AI timelines are existentially devastating in a way that nobody seems to want to discuss - people who work at frontier AI labs will say with a straight face that they think the singularity is happening within 5 years and there has been almost no discussion that i've seen of what that means in human terms. it's just too overwhelming and insane. on some level believing this + the standard AI doomer arguments (which, god help me, i continue to find compelling) means believing that every living thing on this planet has a terminal illness it's about to die of any year now, and nobody has a plausible plan for curing it, because of fucking capitalism or geopolitics or some shit. the main upshot for me personally is that i've completely stopped worrying about climate change because i viscerally don't believe it will happen fast enough to outspeed the singularity i mentioned before that it used to be sort of a meme in the rationality community that AI timelines meant it didn't make sense to save for retirement, but at this point my personal timeline is now short enough that no plan that takes longer than... 2 years to come to fruition viscerally feels like it makes any sense anymore. i am no longer sure it makes sense to work towards buying a house or having kids. the entire ordinary script for how to have a fulfilling ordinary life doesn't make sense to me anymore the singularity is a black hole in my mind that sucks the meaning out of things and every time a new state-of-the-art model releases it grows bigger and bigger and bigger. i don't know how to talk about any of this both because i don't want to get into tiresome arguments and because i don't want to freak out anyone who might be vulnerable to freaking out about this. insofar as i've spent much of the last several years exorcising the ghosts that haunt me, this is the last and biggest and worst one, this is the final boss, and i have no idea what to do about it, i have no idea if i'm strong enough the main way i cope with the black hole is by taking things one day at a time and focusing on what's in front of me which i suppose is as good a way to pass the time as any. when i go for a walk through the neighborhood and look at trees in the sunlight everything looks perfectly ordinary and sane and i can forget. there are a lot of ways to keep busy. it's only when i've run out of things to do and i have to be alone with my mind and it runs out of ordinary life problems to fret about that the black hole creeps back into my awareness, growing and growing and growing, a mental tumor i have not figured out how to excise

@scheminglunatic Join the gitter

Hm, Futhark is a cool language (has a lot of things I want in terms of intersecting numerical computing with functional programming); but what will the F#ers think of them using the hedgehog?

@scheminglunatic just the right time try futhark-lang.org

tiem to catch up on AoC

@materzynska Incredibly important

1/ 🎨✨ If an AI has never seen art before, can it develop artistic capabilities? Our team from MIT, Northeastern and ShanghaiTech University analyzes this question in our Art-Free Diffusion paper.

@nearcyan 🥲

the pre-slop AI art era was so beautiful imo the entire community consisted of a few hundred people (pics here were >6mo before stable diffusion or chatgpt) the ephemerality of such periods begets homesickness, but you can never return to it, no matter how badly you long for it