No Strong Orthogonality From Selection Pressure
TL;DRIf everything goes according to plan, by the end of this post we should have…
Category: Opinion
Computation in Superposition: Two Handcrafted Models
Many interpretability researchers (ourselves included) believe that neural networks store knowledge in superposition—that is, networks…
Research Sabotage in ML Codebases
One of the main hopes for AI safety is using AIs to automate AI safety…
Research Sabotage in ML Codebases
One of the main hopes for AI safety is using AIs to automate AI safety…
The fall of the theorem economy (David Bessis)
I found this post from mathematician David Bessis very interesting. It explains that while AI…
Probe-Based Data Attribution: Surfacing and Mitigating Undesirable Behaviors in LLM Post-Training
IntroductionResearch by Frank Xiao (SPAR mentee) and Santiago Aranguri (Goodfire).Post-training can introduce undesired side effects…
Book review: The Infinity Machine
Book review: The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence, by Sebastian…
Poisoning Fine-tuning Datasets of Constitutional Classifiers
The primary contributors to this work are Chase Bowers, Faizan Ali, John Hughes, Jerry Wei,…
Lorxus Does Budget Inkhaven Again: 04/22~04/28
I'm doing Budget Inkhaven again! (I didn't realize last time that "Halfhaven" also meant specifically…
AGI is Probably Inevitable: A Model of Societal Ruptures
Introduction I posted a quick take a few days back claiming AI moratoriums won't work,…