Logan Riggs

Deep models reveal better strategies for superposition

by Bartosz Rzepkowski, Philipp Alexander Kreer, tdooms, and Logan Riggs

1. Introduction The current, incredible performance of AI models is closely related to their compression capabilities (Language Modeling Is Compression (Delétang et al., 2023); Compression Represents Intelligence Linearly (Huang et al., 2024)). This compression is imposed on them by the architectural choices made by engineers. For example, GPT-2 had a...

Sep 2823

Mech Interp is a Verifiable Task If...

If we think parts of MLP0-MLP3 are computing [a sorting algorithm], we can replace those parts with [a sorting algorithm] and check reconstruction loss.[1] However, reconstruction loss is not enough. Suppose we replace MLP0 with two things: 1. Its mean activation - simple, but poor reconstruction 2. MLP0 - perfect...

Sep 2143

If you're interpreting <1B parameter models, you should use a tensor transformer

To all my fellow researchers doing SLT, computational mechanics, one of ARC's programs, natural abstractions/condensation, proofs on NNs (or any interp on small models), this is for you. Tensor transformers (ie replacing your MLPs & attention with bilinear variants) are performant[1] and allow you to deploy the full power of...

Sep 254

How has publishing your research on LW or X been helpful to you?

I work at CBAI, and we might push to get fellows to publish results on LW or X. I'd like to give more than just my own anecdotes. So please comment your experience here (or in this google form if you don't want a public comment). My Experience Better Communication...

Jul 1722

When Are Two Networks the Same? Tensor Similarity for Mechanistic Interpretability

We've found a method that tells you: * How functionally similar two neural networks are across ALL inputs, * Computed solely from the weights (i.e. no data), * Using a principled generalization of cosine similarity. There's only one catch: you have to use a tensor network. We've already shown that...

May 2945

Black Boxes for Low-Stakes, Interpretable AI for High-Stakes

If we have models that are 10x less efficient but completely interpretable[1],this would be a multi-billion dollar industry. If you just needed to train your bio-model 10x longer to [reverse-engineer human bio-markers for dementia], then you can now sell your product. Task-specific models are much, much smaller than general SOTA...

May 2820

Write Cause You Have Something to Say

The ones who are most successful at writeathons (Inkhaven, NaNoWriMo) are those with an overhang of things to say, usually in the form of: * draft posts * daydreams When Scott Alexander said: > Whenever I see a new person who blogs every day, it's very rare that that never...

May 837

Logan Riggs

Logan Riggs

Hospitalization: A Review

Seeking Power is Often Convergently Instrumental in MDPs

Sparse Autoencoders Find Highly Interpretable Directions in Language Models

A descriptive, not prescriptive, overview of current AI Alignment Research

Logan Riggs

Hospitalization: A Review

Seeking Power is Often Convergently Instrumental in MDPs

Sparse Autoencoders Find Highly Interpretable Directions in Language Models

A descriptive, not prescriptive, overview of current AI Alignment Research

Deep models reveal better strategies for superposition

Mech Interp is a Verifiable Task If...

If you're interpreting <1B parameter models, you should use a tensor transformer

How has publishing your research on LW or X been helpful to you?

When Are Two Networks the Same? Tensor Similarity for Mechanistic Interpretability

Black Boxes for Low-Stakes, Interpretable AI for High-Stakes

Write Cause You Have Something to Say