The biggest event in 2025 for me is entering the field of AI safety research. In particular, I work with collaborators at Geodesic Research on designing a suite of health metrics to evaluate the "pathologies" in the chain-of-thought reasoning of Large Language Models, such as post-hoc, internalized, and encoded reasoning. I've also started reading the Philosophical Investigations by Ludwig Wittgenstein in my leisure time. To my pleasant surprise, Wittgenstein investigated these same phenomena in human language in his famous work. Given the speed of LLM development today and the urgent need for upholding monitorability of chain-of-thought reasoning (see this recent post from OpenAI), I find his work enlightening for drawing parallels between... (read 985 more words →)
I would appreciate a lecture on how causality is used in mechanistic interpretability, particularly for understanding the faithfulness of Chain-of-Thought reasoning. This intersection is crucial for AI safety as it addresses whether LLMs actually follow the reasoning steps they generate or if their explanations are post-hoc rationalizations. The field has developed sophisticated causal intervention methods to probe the relationship between intermediate reasoning steps and final outputs.
Key Papers: