Alek Westover

Right now I'm working on AI safety at Redwood Research.
Before joining Redwood I did theoretical computer science research at MIT
(data structures, online algorithms, and algorithmic graph theory).

Alek's CV

Contact

Email: alek dot westover at gmail dot com

Blog

I occasionally put unpolished thoughts about life (and sometimes math) on my blog here.
I write about AI safety on LessWrong.

LessWrong Posts

• [Paper] Distillation for Incrimination and Distillation for Capabilities
• Research Note: Filtering Subversion-Relevant Information From Pretraining Data Is Feasible
• Latent reasoning architectures would likely undermine CoT, our strongest oversight tool
• Deep recurrent models are less robustly CoT-monitorable than normal CoT models in a toy setting
• Astra is much better at reasoning with filler tokens than previous models
• An operationalization of opaque serial depth
• Proposal for tracking the effects of architecture on monitorability
• The distillation double bind: Distilling misaligned models either transfers misalignment or it doesn't
• Advice for making robust-to-training model organisms
• Why does off-model SFT degrade capabilities?
• Incriminating misaligned AI models via distillation
• Sleeper Agent Backdoor Results Are Messy
• How do LLMs generalize when we do training that is intuitively compatible with two off-distribution behaviors?
• Model organisms researchers should check whether high LRs defeat their model organisms
• Should AI Developers Remove Discussion of AI Misalignment from AI Training Data?
• What training data should developers filter to reduce risk from misaligned AI? An initial narrow proposal
• What is (and isn’t) gained by avoiding architectures with high opaque serial depth?
• Malign initializations are more robust when the model can think better in the reasoning language than in the output language
• A list of existing alignment approaches
• How to reduce capability degradation from off-model SFT
• An Empirical Study of Methods for SFTing Opaque Reasoning Models
• Five approaches to evaluating training-based control measures
• How will we do SFT on models with opaque reasoning?
• Three visions for diffuse control
• Theoretical predictions on the sample efficiency of training policies and activation monitors
• Four Downsides of Training Policies Online
• Methodological considerations in making malign initializations for control research
• Notes on Software-Based Compute-Usage Verification
• Why I think AI will go poorly for humanity
• Safe Distillation With a Powerful Untrusted AI

Publications

• "Distillation for Incrimination and Distillation for Capabilities"
  Sebastian Prasanna, Jacqueline Tay, Alek Westover

• "Lower Bounds for Linear Hashing via Arithmetic Kakeya"
  Ainesh Bakshi, Alex Conway, Hanna Komlós, William Kuszmaul, Alek Westover

• "Listing 6-Cycles in Sparse Graphs" (ITCS'25)
  Virginia Vassilevska Williams, Alek Westover

• "The Diamond Test: A Novel Direct Sum / Affinity Test" (ITCS'25)
  Alek Westover, Edward Yu, Kai Zheng

• "When to Give Up on a Parallel Implementation" (ITCS'25)
  Nathan S. Sheffield, Alek Westover

• "Interrater Reliability Estimation via Maximum Likelihood for Gwet’s Chance Agreement Model." (OJoS'24)
  Westover, Alek M., Tara M. Westover, and M. Brandon Westover.

• “A Nearly Quadratic Improvement for Memory Reallocation” (SPAA'24) (Slides)
  Martin Farach-Colton, William Kuszmaul, Nathan Sheffield, and Alek Westover

• “Scheduling Jobs with Work-Inefficient Parallel Solutions” (SPAA'24) (Slides)
  William Kuszmaul and Alek Westover

• “Complexity of Multiple-Hamiltonicity in Graphs of Bounded Degree”
  Brian Liu, Nathan S. Sheffield, Alek Westover

• “On the Relationship Between Several Variants of the Linear Hashing Conjecture”
  Alek Westover

• “The Variable-Processor Cup Game” (ITCS'21) (Slides)
  William Kuszmaul and Alek Westover

• “Cache-Efficient Parallel-Partition Algorithms using Exclusive-Read-and-Write Memory” (SPAA'20) (Animation) (GitHub)
  William Kuszmaul and Alek Westover