Ameya Panchal

Now

CS, physics & AI at Penn State.Independent interpretability research.

I am in the process of publishing a neutron-star paper, and am transitioning into AI safety research in mechanistic interpretability. Currently I am working to reproduce the early-layer degradation result from the Global Workspace paper (Gurnee et al., 2026).

Current experiment · Does Anthropic's global workspace result replicate?

Writing03

Where the Error Hides: The Isotropy Bias in Neutron-Star Pulse Profiles

A Monte Carlo scattering atmosphere puts a number on a standard modeling shortcut: +0.14 in pulsed fraction. Then geometry and spin make the number vanish — and the error doesn’t disappear, it relocates.

Aug 4, 2026

7 min

Every Method Has a Blind Spot: Reading GPT-2’s IOI Circuit Three Ways

DLA finds the name movers, patching finds what feeds them, path patching finds the wire between — and each one is blind to exactly what the next reveals.

Jul 22, 2026

13 min

An Induction Head in Disguise: Chasing Grammar in a Character-Level Transformer

A copying head with the OV half of a bracket-completion mechanism, but a QK circuit that never learned to route it.

Jul 10, 2026

7 min

Notes03

Where the Error Hides: The Isotropy Bias in Neutron-Star Pulse Profiles

Every Method Has a Blind Spot: Reading GPT-2’s IOI Circuit Three Ways

An Induction Head in Disguise: Chasing Grammar in a Character-Level Transformer

Better than the last one

A statue with reflexes

A dead-ReLU absorbing state