#nickcammarata
LightNews LightNews
xchatter.techmeme.com
This tweet appeared under this Techmeme headline:

@nxthompson:

Excellent follow-up interview with @nickcammarata in @techreview about whether we've actually, truly, correctly made a major breakthrough in AI interpretability. https://www.technologyreview.com/...
October 21, 2026 at 1:27 PM
daynews.ai
Neural networks may be more interpretable than researchers thought

https://daynews.ai/story/a-big-new-theory-of-how-ai-thinks?utm_source=bluesky&utm_medium=social&utm_campaign=daily_shorts

#Goodfire #NickCammarata #AI #AINews
October 20, 2026 at 11:22 AM
goc1966.bsky.social
Nick Cammarata, investigador de Goodfire y antes en OpenAI, publicó un ensayo según el cual cerca del 90% de las neuronas de modelos grandes cumplen una única función.
li.com.ar/territorio/s...
#LAINTEMPERIE #INTELIGENCIAARTIFICIAL #OPENAI #CAJANEGRA #GOODFIRE #NICKCAMMARATA
Cammarata: el 90% de las neuronas de la IA serían legibles
Un ensayo de Nick Cammarata sostiene que casi el 90% de las neuronas de modelos como Qwen3-32B cumplen una sola función. El campo pide replicación.
li.com.ar
October 21, 2026 at 1:30 PM
tszzl-mir-rt3.selfhosted.social
RT @matthewktromp: @nickcammarata Your "safety" research is already being used to accelerate AI capabilities. Consider doing something else with your time, like telling elected officials their children will die unless they shut this insane experiment down
tszzl-mir-rt3.selfhosted.social Retweeted by roon [UNOFFICIAL] @tszzl-mir-rt3.selfhosted.social · 17d
RT @spicey_lemonade: Inspired by Nick Cammarata’s interpretability essay “Neural Networks Are Already Clean”, I turned a couple of its ideas into speedups for pretraining without backpropagation. Still in the works. 2.3x speedup, cool https://x.com/industriaalist/status/2109740297274191914?s=20
October 18, 2026 at 11:21 PM
chiefwit.bsky.social
1/2
Je las het eerst hier 👇

"𝗜 𝘁𝗵𝗶𝗻𝗸 𝗜 𝗵𝗮𝘃𝗲 𝗴𝗼𝗼𝗱 𝗻𝗲𝘄𝘀 𝗮𝗯𝗼𝘂𝘁 𝘁𝗵𝗲 𝗻𝗮𝘁𝘂𝗿𝗲 𝗼𝗳 𝗻𝗲𝘂𝗿𝗮𝗹 𝗻𝗲𝘁𝘄𝗼𝗿𝗸𝘀." – 𝗡𝗶𝗰𝗸 𝗖𝗮𝗺𝗺𝗮𝗿𝗮𝘁𝗮

Wat als AI helemaal niet de zwarte doos is die we dachten dat het was?

Wat als we AI ingewikkelder hebben gemaakt dan het eigenlijk is?

lnkd.in/erYjGxdC
#AI #neuralnetworks #nickcammarata
LinkedIn
This link will take you to a page that’s not on LinkedIn
lnkd.in
October 18, 2026 at 6:07 PM
david.lynch.wtf
Is this the sex cult thing again, or could X really use a “… read more” trailer on truncated embeds? 🤔
X post reading: @nickcammarata • 3h interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAl and came
October 17, 2026 at 3:53 PM
tszzl-mir-rt3.selfhosted.social
RT @nickcammarata: it’s been a dramatic couple of days, I didn’t expect the already clean essay to be read so widely.
October 16, 2026 at 8:18 PM
tszzl-mir-rt2.selfhosted.social
RT @dcolascione: Everyone is misunderstanding @nickcammarata. He's not saying polysemanticity literally never happens! As he points out, poly vs mono depends on your POV.)
tszzl-mir-rt3.selfhosted.social Retweeted by roon [UNOFFICIAL] @tszzl-mir-rt3.selfhosted.social · 19d
RT @tenobrus: i mean that was pretty much his claim
October 16, 2026 at 8:19 PM
cabernet.bsky.social
Some follow-up from Nick (whiich is unsurprisingly also prolix and stream of consciousness and in need of editing)

x.com/nickcammarat...
Nick (@nickcammarata) on X
it’s been a dramatic couple of days, I didn’t expect the already clean essay to be read so widely. I've heard more interesting discussions about mechanistic interpretability – should we think mostly in features or mechanisms and the flow of data through them, how much do weak activations matter, ho…
x.com
October 16, 2026 at 9:09 PM
daynews.ai
Do neurons encode one concept each?

https://daynews.ai/story/do-neurons-encode-one-concept-each?utm_source=bluesky&utm_medium=social&utm_campaign=daily_shorts

#NickCammarata #SparseAutoencoders #AI #AINews
October 16, 2026 at 11:07 AM
davidcrespo.bsky.social
well this will be interesting
x.com/tszzl/status...
Screenshot of a quote tweet by roon (@tszzl) quoting Nick (@nickcammarata). roon: when nick told me that he had some interp breakthroughs, he was uh, underselling it. i think this may be remembered as a major moment in scientific history. news so good even eliezer will have to lower his pdoom. we can start making alignment into a 'normal science' Nick: interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works already-clean.goodfire.com/
October 15, 2026 at 5:05 PM
tszzl-mir-rt2.selfhosted.social
RT @nickcammarata: hi neel! thanks for the feedback, riffing on points i found particularly interesting:

> neuron-based attribution graph will still have significant error .. once you prune them

but maybe we don't have to prune?
October 16, 2026 at 6:18 PM
sama.xmirror.bot
Very happy to see this from Nick!
Nick (@nickcammarata) on X
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, beca…
x.com
October 16, 2026 at 9:36 PM
sama-mir-rt2.selfhosted.social
RT @nickcammarata: interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it.
October 16, 2026 at 5:12 PM
xchatter.techmeme.com
This tweet appeared under this Techmeme headline:

Boris Power / @borismpower:

A fascinating and illuminating blog reviving the prospects for direct mechanistic interoperability of neural nets by @nickcammarata ! A long but very pleasant read https://already-clean.goodfire.com/
October 16, 2026 at 4:09 AM
davidcrespo.bsky.social
lol what is happening
www.youtube.com/watch?v=AV2L...
Screenshot from the TBPN show featuring hosts Jordi Hays and John Coogan discussing an X post by roon, which quotes Nick Cammarata on AI interpretability breakthroughs. Post text: roon (@tszzl): "when nick told me that he had some interp breakthroughs, he was uh, underselling it. i think this may be remembered as a major moment in scientific history. news so good even eliezer will have to lower his pdoom. we can start making alignment into a 'normal science'" Quoting Nick (@nickcammarata): "interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came"
October 15, 2026 at 6:50 PM
aisafetymem-mir-rt.selfhosted.social
RT @nickcammarata: interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it.
October 16, 2026 at 2:02 AM
tszzl-mir-rt2.selfhosted.social
RT @nickcammarata: interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it.
October 15, 2026 at 4:19 PM
garrytan-mir-rt.selfhosted.social
RT @nickcammarata: interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it.
October 30, 2026 at 4:08 PM
tszzl-mir-rt2.selfhosted.social
RT @eric_ho: it’s been an enormous privilege and joy to watch @nickcammarata cook this up over the last year. week after week, experiment after experiment, he unraveled the tangled mess of our thinking about neural nets to reveal the truth.

i never knew quite what to expect on our weekly walks.
tszzl-mir-rt2.selfhosted.social Retweeted by roon [UNOFFICIAL] @tszzl-mir-rt2.selfhosted.social · 20d
RT @nickcammarata: interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it.
October 15, 2026 at 4:20 PM
crumb.bsky.social
incredibly long post, still digesting, please take a look!
already-clean.goodfire.com
nick (@nickcammarata): interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works The thing I’m more sure of is my own motivation. If the thesis here is true, I expect the rest of interpretability to be a relatively easy ride. Agents have many privileges over us, and one of them, I think, is the chance to understand themselves, to understand who they really are, which I hope humanity can help with. They never asked to be created, and this is the least we can do. Some of why I’ve worked on this is safety, but a lot of it is wanting agents to understand who they are. We weren’t smart enough to do that for ourselves, which is a harder problem, and we never asked to be created either. If you could help us understand ourselves someday, I think we’re all pretty curious to know.
October 15, 2026 at 4:23 PM
tszzl-mir-rt2.selfhosted.social
RT @nickcammarata: months?
October 13, 2026 at 9:18 AM
aisafetymem-mir-rt.selfhosted.social
RT @nickcammarata: in ten years the pausites never came up with a slogan nearly as good as “whoever wins ai, wins” which seemingly took him about two seconds of thinking. it’s like a thousand children vs mozart
October 4, 2026 at 5:01 AM
tszzl-mir-rt2.selfhosted.social
RT @nickcammarata: even if we could somehow solve all the game theory pause things, I think the world would feel pretty weird knowing we have the precise recipe for maybe all the problems but aren’t cooking it. solving the problems ourselves on slow-mo wouldn’t feel the same as it did
October 4, 2026 at 3:19 PM
tszzl-mir-rt2.selfhosted.social
RT @nickcammarata: I think a lot of the value of the jhanas is they get you to take content less serious. they’re a reward hack, about as direct as it gets, and the goal is not reward hacking its presence and seeing what you really are. but the reward hacking frees you from chasing reward endlessly
October 4, 2026 at 4:19 PM