AI Safety

The Remarkable Robustness of LLMs: Stages of Inference?

We find that deleting and swapping interventions retain 72-95% of the original model’s prediction accuracy without fine-tuning, and hypothesize the existence of four universal stages of inference across eight different models.

Mechanistic Interpretability for Progress Towards Quantitative AI Safety

MIT Master’s thesis studying mechanistic interpretability as a path toward quantitative AI safety.

Opening the AI Black Box: Distilling Machine-Learned Algorithms into Code

We present MIPS, a novel method for program synthesis based on automated mechanistic interpretability of neural networks trained to perform the desired task, auto-distilling the learned algorithm into Python code.