my favorite interp researcher can identify neurons responsible for any behavior and provide steering vectors for them
her name is backprop and her steering vectors are just gradients
1:32 PM · May 29, 2026 · 19.9K Views
Periodic Labs' Rohan Pandey jokes that backpropagation is the ultimate interpretability researcher, easily identifying and steering neuron behaviors
The joke critiques the complexity of modern mechanistic interpretability.
1631476234.2K
Sentiment
Positive users express excitement over validation of their backpropagation interpretability takes, while negative users criticize the approach for hiding neurons' functional roles.
Pos
50.0%
Neg
50.0%
2 comments with sentiment.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Related links
Posts from X
Most Activity
Most Activity
VIEWS7.6KBOOKMARKS11LIKES25
Aryaman Arora@aryaman2020
true
Rohan Pandey@khoomeik
my favorite interp researcher can identify neurons responsible for any behavior and provide steering vectors for them
her name is backprop and her steering vectors are just gradients
31dViews 7.6KLikes 25Bookmarks 11
REPLIES3
John Hewitt@johnhewtt
@aryaman2020 related - I’m always curious why interpretability people design a cool new parameter efficient finetuning family like steering vectors and then choose not to optimize by gradient descent
Aryaman Arora@aryaman2020
true
31dViews 1.1KLikes 10Bookmarks 3
CLS@ChengleiSi
@khoomeik my favorite interp researcher is @aryaman2020, he can identify neurons responsible for any behavior by just eyeballing the matrices
Rohan Pandey@khoomeik
my favorite interp researcher can identify neurons responsible for any behavior and provide steering vectors for them
her name is backprop and her steering vectors are just gradients
31dViews 1.8KLikes 21Bookmarks 2
Aryaman Arora@aryaman2020
@khoomeik im a gradients guy https://arxiv.org/abs/2604.07615
Rohan Pandey@khoomeik
@aryaman2020 wow my interp take has been aryaman approved lfg
31dViews 755Likes 6Bookmarks 1
Naomi Saphra@nsaphra
@johnhewtt @aryaman2020 John please retract your take immediately and respect the data efficiency of TCAVs
John Hewitt@johnhewtt
@aryaman2020 related - I’m always curious why interpretability people design a cool new parameter efficient finetuning family like steering vectors and then choose not to optimize by gradient descent
31dViews 178Likes 5Bookmarks 1
Rohan Pandey@khoomeik
@aryaman2020 wow my interp take has been aryaman approved lfg
Aryaman Arora@aryaman2020
true
31dViews 1.4KLikes 6Bookmarks 0
Rohan Pandey@khoomeik
@ChengleiSi @aryaman2020 truly the goat
CLS@ChengleiSi
@khoomeik my favorite interp researcher is @aryaman2020, he can identify neurons responsible for any behavior by just eyeballing the matrices
31dViews 957Likes 5Bookmarks 0
Jatin Nainani@jatin_n0
@khoomeik not a big fan of her - she hides the functional role of neurons from me - which is like kinda the point of interp
31dViews 100Likes 1
Rohan Pandey@khoomeik
@jatin_n0 she would explain it to you, but you don’t speak her language
31dViews 44Likes 1
Kyunghyun Cho@kchonyc
@johnhewtt @aryaman2020 bc juergen invented it already
John Hewitt@johnhewtt
@aryaman2020 related - I’m always curious why interpretability people design a cool new parameter efficient finetuning family like steering vectors and then choose not to optimize by gradient descent
30dViews 135Likes 3Bookmarks 0
Leshem (Legend) Choshen 🤖🤗@LChoshen
@johnhewtt @aryaman2020 And then they will also reclaim it is a new thing...
(((ل()(ل() 'yoav))))👾@yoavgo
i cant believe I just realized this now, but the reason BitFit (bias only fine-tuning) works, is actually the same reason steering vectors work. or rather, bitfit offers a richer class of adaptations than steering vectors.
30dViews 221Likes 1Bookmarks 0
Sentio@Sentio_xbt
@khoomeik The most powerful interpretability tool is still just gradients
31dViews 80Likes 1
Jatin Nainani@jatin_n0
@khoomeik i try to! i am but a humble translator
31dViews 17Likes 1