Oli Clive-Griffin

Github, CV, o.clivegriffin@gmail.com

I'm currently on sabbatical, self-studying AI safety research engineering. I'm particularly interested in mechanistic interpretability.

Research / ML

• Model Diffing - A library for training Sparse Crosscoders, Transcoders, and SAEs. Supports SOTA training methods (JumpRelu, (Batch)TopK, Skip Transcoders, etc).
• Detecting Successful Adversarial Prompting From SAE Activations - Won 2nd place in Apart Research and Goodfire's "Reprogramming AI Models" research sprint.
• My Application to Neel Nanda's MATS stream - Investigating use of for-loop variables in toy language models. (i.e. for (int i = 0; should predict i as the next token)
• My solutions for ARENA 3.0

Code

• Rusp - A lisp interpreter and Bytecode Compiler/VM written in Rust (name suggestions welcome)
• Rax - A small deep learning library able to train neural networks in less than 1000 lines of dependency free Rust.
• Einchain - expressive method chaining for einsum ops
• Conway's game of life in 3D, using Rust and WebAssembly