InterpretabilityResearch

Toy Models of Superposition

Sep 14, 2022
Read Paper

Abstract

In this paper, we use toy models — small ReLU networks trained on synthetic data with sparse input features — to investigate how and when models represent more features than they have dimensions. We call this phenomenon superposition. When features are sparse, superposition allows compression beyond what a linear model would do, at the cost of "interference" that requires nonlinear filtering.

Related content

Learning more about Claude's mathematical capabilities

An unreleased research version of Claude has made strides on a problem related to the Riemann hypothesis. It improved a longstanding lower bound for the fraction of zeros of the Riemann zeta function that satisfy the hypothesis, increasing it from 41.6% to 67.2%.

Read more

Discovering cryptographic weaknesses with Claude

cryptographic algorithms. The first attack significantly weakens HAWK, a digital signature scheme that was built for a future world where quantum computers are able to break existing standards. The second identifies a new way to attack round-reduced AES, the most widely used symmetric cipher.

Read more

Project Pilot: Can AI control a drone?

Working with Andon Labs, we’ve developed a new series of evaluations that assess AI models’ ability to use a flying drone, culminating in a new benchmark: Drone-Bench.

Read more