roberthuben Goto Github PK

activation_correlations

Simulates activations of neurons and features, and computes expected correlations.

emergent-world-representations-othello

A mechanistic interpretability study invvestigating a sequential model trained to play the board game Othello

factor_orbits

Creates animations of N dots spinning in an orbit that displays the factors of N

ffn_via_attention

Implements the components of a transformer (including feedforward networks) entirely via attention heads

imagenet_adversarial_noise

Adds adversarial noise to images to fool imagenet.

incorrectness_cascade

Generate and analyze data about whether incorrect answers from GPT 3.5 results in more incorrect answers

incorrectness_cascade_b

othellogpt_sparse_autoencoders

procgenaisc

sparse-autoencoders-figures

Data and code for making figures for the Sparse Autoencoders paper

sparse_coding

Work on sparse coding, replicating and extending the sparse coding approach to taking transformer features out of superposition.

tree_point_clouds

Methods for identifying individual trees from point cloud data

word-embeddings

Testing properties of GPT-J word embeddings in order to solve the SolidGoldMagikarp mystery