This is your work, valued

London

Callum McDougall

Elite
@callummcdougall

Interpretability Researcher at Google DeepMind. ex-director of the ARENA program.

ARENA_3.0. Jupyter Notebook

1.2k

sae_vis. Create feature-centric and prompt-centric visualizations for sparse autoencoders (like those from Anthropic's published research).

262

ARENA_2.0. Resources for skilling up in AI alignment research engineering. Covers basics of deep learning, mechanistic interpretability, and RL.

246

computational-thread-art. Code that renders an image as a series of lines connecting pins around a circular frame (for more detail, see my Medium page).

140

sae_visualizer. HTML

31

sae-exercises-mats. Python

26

TransformerLens-intro. HTML

20

path_patching. Implementation of path patching & activation patching (will eventually add to TransformerLens).

15

computational-thread-art-v2. Updated version of my computational thread art algorithm, which also includes code for generating full-color images.

10

SERI-MATS-2023-Streamlit-pages. Repo for hosting Streamlit pages for my 2023 SERI MATS project with Arthur Conmy (mentored by Neel Nanda).

9

CircuitsVis. Mechanistic Interpretability Visualizations using React

8

jupyter-to-anki. Python functions to convert Jupyter Notebooks into Anki cards.

7

eindex. My interpretation of what einops indexing would look like (created to work on during my SERI MATS project).

6

arena-heroku. HTML

6

anki_templates. Contains templates for Anki cards, and code to auto-generate Anki cards (coming soon)

5

computational-thread-art-v3. Jupyter Notebook

5

arena-pragmatic-interp. Jupyter Notebook

5

SERIMATS_app. Application for Neel's SERI MATS stream, 2023.

3

arena-v1. HTML

3

plotly-widgets. Jupyter Notebook

2

SAE_Bench_Template. Python

2

Fundamentals. HTML

1

arena-v1-ldn-exercises-new. HTML

1

copy-suppression. HTML

1