Surprising facts about the slave trade
1. The obstacle to abolition was not the economic system, but an industry lobby.I had always imagined the British abolitionist movement to be a broad battle…
Latest
Request for mentors to join the Sentient Future Project Incubator
I helped Sentient Futures run the first iteration of their Project Incubator. The program brought together around 120 people, split between mentors and mentees, to work…
Intelligence (Artificial) in a Fallen World
Cross-posted from babbo.dev/articles/ai-risk. To experience the piece in its intended form, please visit there.A Kishōtenketsu on AI RiskFaces*, Madge Gill, 1920–1960. Newham Archives and Local Studies…
Podcasts: AI and You, and Me
I was on Peter Scott’s podcast, Artificial Intelligence and You and our discussion apparently took two episodes! (Quotes from Peter about what we discussed)June 15: https://aiandyou.net/e/313-guest-katja-grace-ai-impact-researcher-part-1/We…
Chorus-Reinterpretation Country Songs
Our family is on vacation in North Carolina for a week, spending some time at a pool, and they're playing a (weirdly short) loop of music.…
Exploring Generalization in NLA's
Recently, I was reading anthropic's paper on NLA's[1] and for a person who works on steering, it was an interesting and thought-provoking paper. In this post…
fab: how to do (alignment) research at scale
Over the last month, I've been working on a project I'm calling fab.[1]Nominally, it's an interface that enables a human researcher to make sense of research…
Eval-Awareness Steering detects the Test, Not the Sabotage
Produced as part of independent researchHuge thanks to Apollo Research (org) for open-sourcing the deception-detection harness which proved to be foundational in this work. Prior work…
Dev log: implementing new syntax in GHC
In September 2024, I started working on a patch to GHC. The story starts slightly before that. In July 2024 I wondered Could UndecidableInstances be per-instance?.…
Exploration: fine-tuning with parameter decomposition
TL;DR: We can destroy a 67M-parameter language model's ability to predict German text by fine-tuning a single number: the scalar prefactor on one German-related rank-1 parameter…