Toward Agents + Robustness, Visualization, and Adversaries

Morning Outline: Toward agents

• Prediction as Understanding (next-token prediction and emergent intelligence)
• Emergent Agency (tools as tokens, persistent memory, and iterative loops)
• Model-Centric Optimization (alignment via prompt engineering and reinforcement learning)
• slides

Afternoon Outline: Robustness and Visualization and Adversaries

• Jagged Intelligence and Robustness (advanced capabilities vs. simple failures; adversarial attacks)
• Applications in Investigations (matching hotel room photos to identify crime locations)
• CLIP and Foundation Models (shared high-dimensional vector spaces; embedding limitations)
• slides

Lab Activity

The afternoon activity is Visual Adversarial Attacks.