Udari Madhushani Sehwag

LLMs can't reliably predict scientific experiment outcomes, and more worryingly, they have no idea when they're wrong, unlike human experts whose accuracy skyrockets when they feel confident.

Safety-aligned LLMs are accidentally handicapping cyber defenders, refusing to help with critical tasks like malware analysis and system hardening simply because the requests sound too much like attacks.