Welcome!

I’m currently at Anthropic working on evaluating misalignment risk. I previously worked at Conjecture and at Google in a team collaborating with DeepMind.

• [2024-04-01] Scaling calculator
• [2024-04-01] Sparse autoencoders depend too much on theories
• [2024-03-15] Breaking down Situational Awareness
• [2024-02-24] How well do truth probes generalise?
• [2024-02-21] Jailbreaking GPT-4 with the tool API
• [2024-02-15] Conway’s law and coding assistants
• [2022-12-31] AI x-risk model (perpetual draft)
• [2022-10-18] Distilled Representations Research Agenda
• [2022-10-12] Preventing steganography in latent vectors
• [2022-09-25] Feature visualisation
• [2021-03-13] vaxtldr.uk: England vaccination progress TL;DR
• [2020-02-17] Introducing KIPA: Distributed Key to IP Address translation
• [2018-11-21] Answering Visual What-If Questions: From Actions to Predicted Scene Descriptions

I often write to think things through, but I then don’t robustly critique or tidy my writing, resulting in writing that conveys something with low confidence. But I still like to have them linked:

• [2024-03-17] In research, optimise for bitrate
• [2024-02-28] Correlated failures and coordinated misalignment
• [2024-02-28] Concerns with RSPs
• [2024-01-07] Evals are as hard as product development
• [2023-05-13] Trust