Bruce W. Lee

I work on reducing risks from misaligned agents with Tomek Korbak. Previously, I worked with Alex Turner and Alex Cloud on capability scoping.

Before moving to the US, I served in the Republic of Korea Marines, where I operated helicoptersKMEP helicopter operations. Now, I'm on Penn's men's lightweight rowing teamPenn Lightweight Rowing.

Sometimes I have thoughts that don't fit into a paper or a post elsewhere. I put those here. If you only read one thing, I recommend Salient Directions in AI Control.

Writings
Blog Posts
Mar 2026 Salient Directions in AI Control 19 min read Feb 2026 What Would It Take to Trust AI for AI Safety? 5 min read Dec 2025 Bitter Lessons from "Distillation Robustifies Unlearning" 10 min read Aug 2025 Neural Networks, Strange Attractors, and Orderliness in Chaos 10 min read Nov 2024 On Getting Started in Research 3 min read Sep 2024 Mechanistically Programming a Language Model's Behavior 13 min read
Home Blurbs
Short Technical Blurbs
Feb 2026 Why Use "Instructed" Side Tasks for AI Control? 2 min Dec 2025 Loss for an example is just the negative of the logprob that the model assigned to the true class? 2 min Nov 2025 Error Bars as Degrees of Belief 6 min Nov 2025 Reasoning about Neural Network Training with Bias-Variance Tradeoff 6 min Oct 2025 Information Theory and Logistic Regression Both Arrive at Cross-Entropy 7 min Oct 2025 Equivariance and Invariance Explain So Much of Deep Learning 4 min Sep 2025 Why Not Initialize Neural Network Weight to Zero? 3 min
Blog Posts