My research sits at the intersection of AI safety, mechanistic interpretability, and understanding the failure modes of large language models. I focus on confabulations ("hallucinations")—when models produce confident-but-wrong outputs—and methods to anchor outputs to verifiable context. I'm particularly drawn to mechanistic interpretability: tracing how specific neurons, attention heads, and pathways give rise to model capabilities (and failures).

I see language models as more than statistical compression engines—but not as reasoning systems. They exhibit what Ethan Mollick calls "jagged intelligence": superhuman at some tasks, brittle at others that humans find trivial. Karpathy and Demis Hassabis have observed the same pattern—capabilities cluster unpredictably, and fluent outputs mask underlying brittleness. This jagged profile is why interpretability matters: we need to understand where and why systems break.

I'm skeptical that scaling language modeling alone leads to robust reasoning. As Yann LeCun argues, the pretraining paradigm is a dead end for true understanding—these systems master "System 1" thinking but lack the deliberative, causal reasoning of "System 2." This is why I track world models: latent simulators that support planning, counterfactual reasoning, and transfer across domains. Projects like Meta's V-JEPA point toward AI that simulates rather than predicts. I've written more on this in my State of AI 2025 essay—I believe world models are the more interesting frontier.