Aharon Azulay

Self-directed Research | AI Safety & Interpretability

• ARENA (Alignment Research Engineer Accelerator): Comprehensive self-study of foundational interpretability tools, mastered key mechanistic interpretability best practices.
• Extensive study of recent AI safety literature, regularly reading leading research papers in alignment, evaluations, and interpretability.
• Independent research:
  • Triviality Detector with LLMs: Introduced a metric for evaluating model outputs' triviality, demonstrating a novel approach to scalable oversight for frontier model alignment. (Substack article)
  • Obscure Scientific Facts Benchmark for LLMs: Designed to evaluate the factuality of frontier LLMs.
  • Analysis of Chess Grandmaster-level Transformer: Conducted a detailed mechanistic interpretability study of Grandmaster-level chess-playing transformers. Explored fingerprints for generalization vs memorization and implicit planning.
  • Does Competence Increase Stability? A Chess Analogy - Discussed the risks and benefits of open-sourcing AI.

AI Researcher | Lightricks

• Transformer-based Video Editing: Led development of a novel transformer architecture enabling temporally consistent real-time editing on a single V100 GPU [Example output].
• Automatic Color Correction: Deployed RLHF-inspired pipeline enhancing color accuracy in images, validated with user metrics across millions of users.
• Video Temporal Consistency (WACV, 2022): Published a method for improving temporal stability of generative video editing.
• Bokeh Simulation System: Engineered a depth-aware rendering system for realistic blur effects, integrated in flagship apps and used millions of times.
• AI Curriculum Development: Initiated and delivered an internal AI training curriculum covering alignment, interpretability, fine-tuning, reinforcement learning, and agentic systems, enhancing the tech department's (100+ people) alignment with frontier AI developments. [Recording of one lecture about ways to invest test-time compute, predicting the latest trend 5 months before o1 release]

PhD, Machine Learning

Hebrew University (Edmond & Lily Safra Center) | 2016 – 2021

Advisor: Prof. Yair Weiss

Thesis: Why do deep convolutional networks generalize poorly to small image transformations? (JMLR, 2019, 700+ citations). My research focused on neural network robustness, generalization failures, and the connection to adversarial examples.

Relevant Coursework: Advanced theoretical and practical ML, Computational neuroscience, Reinforcement learning, Bayesian inference, Graph theory, and Philosophy of mind.

MSc, Computational Neuroscience

Hebrew University (Edmond & Lily Safra Center) | 2014 – 2016

• Published analysis of C. elegans neural circuits (PLoS Computational Biology, 2016)

BSc, Neuroscience and Amirim Honors Program

Hebrew University | 2011 – 2014

• Magna Cum Laude
• Rector's Prize for Excellent Students

• Interpretability & Alignment: Circuits-based mechanistic interpretability, Sparse Autoencoder (SAE) methods, Probing, Causal Intervention (activation patching, ablation studies), Model Evaluations (Evals), Robustness Analysis, Alignment-focused Transformer Architectures, RLHF and preference learning methods.
• Deep learning: Training and fine-tuning diffusion and autoregressive transformer based models.
• Tools: TransformerLens, SAELens, Vibe coding, PyTorch, Hugging Face, WandB.

Outside of my research, I play chess at a 2000 Elo rating and practice Jiu-Jitsu, which I consider a form of physical chess, at a purple belt level. I enjoy baking sourdough bread and experimenting with fermentation in general. I love music, especially the classical violin repertoire and progressive rock and metal.

Email: aazuleye@gmail.com

Seeking full-time research roles in technical AI safety (interpretability, alignment, evaluations) in Paris, UK, US, or remote.