Ai+ Training

Jagged Intelligence: Limits, Reasoning, and Safety in LLMs

LLMs can ace the bar exam but fail at trivial tasks - a phenomenon known as jagged intelligence. While their peak capabilities are undeniable, their real-world reliability remains brittle.

In this talk, we’ll explore three critical pillars for moving past simple pattern matching toward truly robust AI:

- Mapping the Limits: An empirical look at where and why LLM reasoning breaks down, specifically in out-of-distribution generalization and algorithmic tasks.

- Supercharging Reasoning with RL: How Reinforcement Learning (RL) - driven by structured rewards for partial progress - can unlock genuinely new model behaviors rather than just sharpening existing ones.

- Scaling Safety Systems: Why safety must be a core design constraint. We will dive into next-generation, co-evolving safeguards like WildTeaming (automated adversarial testing) and WildGuard (content moderation).