General support for SaferAI

Support for SaferAI’s technical and governance research and education programs to enable responsible and safe AI.

Train great open-source sparse autoencoders

Find the best settings for SAE training we can, then scale across models

Independent research to improve SAEs (4-6 months)

By rapid iteration on possible alternative architectures & training techniques

Decode Research - Compute for Generating Dashboards & Autointerp

Decoding Logic: Mechanistic Interpretability in LLMs in a research program

Identifying and auditing reasoning circuits in LLMs within Algoverse 2026 using Sparse Autoencoders (SAEs).

Understanding SAE features using Sparse Feature Circuits