Author | Lightnews

jolios torng @tarngerine.bsky.social · 1d

ty @purvigoel.bsky.social @ikauvar.bsky.social joshua batson and adam jermyn for all your work and support on getting this out the door. i learned so much and it's changed how i think about research rigor and communication

7

jolios torng @tarngerine.bsky.social · 1d

Read the full writeup: transformer-circuits.pub/2025/october...

While it’s not a full paper, I’m proud to have been a major contributor. It’s a personal milestone from a non-academic bg, to investigate how LLMs visually reason with talented researchers

Circuits Updates – October 2025

Visual Features Across Modalities: SVG and ASCII Art Reveal Cross-Modal Understanding; Data Point Initialization for Dictionary Models

transformer-circuits.pub

1 2 24

jolios torng @tarngerine.bsky.social · 1d

And these steering gradients are meaningful in transition as well! See how this smile transforms to a :| face before becoming a frown as we steer negatively on the “say smile” motor feature we found above.

1 1 18

jolios torng @tarngerine.bsky.social · 1d

The coolest part is that just like Golden Gate Claude, we can steer with many of these features, transforming a smiley face to wrinkly face, an owl, or an eyeball!

2 2 28

jolios torng @tarngerine.bsky.social · 1d

And yes, we checked that this works for human-made SVGs too! I drew this (imo cuter) dog and found many of the same features as on the Claude-generated dog above (which honestly looks more like a bear!)

1 3 26

jolios torng @tarngerine.bsky.social · 1d

It gets richer as the models improve. Compared to Haiku 3.5, we find a trove of interesting features in the Sonnet 4.5 base model. From various animal body parts, to “motor” neuron features like “say smile” that activate ahead of the ASCII mouth, to features that perceive “size”!

1 5 45

jolios torng @tarngerine.bsky.social · 1d

This understanding is context dependent – the eyes feature light up as soon as you give enough characters to form the top of the head!

Only a single _ is needed, as well as a / \ forehead for the 2nd @ to activate the "eyes" feature above.

1 3 31

jolios torng @tarngerine.bsky.social · 1d

What happens when you turn a designer into an interpretability researcher? They spend hours staring at feature activations in SVG code to see if LLMs actually understand SVGs. It turns out – yes~

We found that semantic concepts transfer across text, ASCII, and SVG:

2 25 130

jolios torng @tarngerine.bsky.social · Sep 22

Funder *

1

jolios torng @tarngerine.bsky.social · Sep 22

FWIW vercel owns svelte too, or rather they’re a major founder and employ rich harris

2 1

jolios torng @tarngerine.bsky.social · Aug 18

made it to Seattle, time to load from this box to this truck then to our house

2

jolios torng @tarngerine.bsky.social · Aug 9

Finally making the move to Seattle! Bye Dallas!

4 11

jolios torng @tarngerine.bsky.social · May 11

I don’t add anyone I don’t know bc ppl do mass add to build legitimacy off their connections

4

jolios torng @tarngerine.bsky.social · Apr 27

you can’t talk about 2x4s and then not show me the shelves!!!

1 1

jolios torng @tarngerine.bsky.social · Apr 23

good job application tbh

1 1

jolios torng @tarngerine.bsky.social · Apr 13

How To Make Lard

2

jolios torng @tarngerine.bsky.social · Apr 5

reinforcement learning is hard and complicated and maybe we just should all agree to stop doing it

1 6

jolios torng @tarngerine.bsky.social · Apr 5

ok so i got Jupyter pilled for months and then my mentor told me they don’t Jupyter anything and just raw dog scripts so… no im on the other side of the fence

jolios torng @tarngerine.bsky.social · Jan 7

i know I should be using jupyter for everything but I can’t get over the mental barrier and I’m just raw dogging the python console in terminal

which is weird bc I’m always screaming about how there needs to be a Jupyter for software design

a lesson somewhere here