Inside the Model

Alignment, Robustness & Oversight

Technical AI safety focuses on a hard truth: models do exactly what they are trained to do, just not what we intended. Self-driving cars mistake children for objects. Image models reproduce illegal content years after training. Language models can be jailbroken, manipulated, or repurposed for cyberattacks.

Topics include:

• Robustness failures and adversarial behavior

• Jailbreaking and misuse-optimized models

• Scalable oversight limits (why humans can’t supervise everything)

• Alignment methods like RLHF, and why they are brittle

• Interpretability and evaluations as early warning systems

Interactive exercises demonstrate how tiny training changes can yield dramatically different behaviors.

Key takeaway: You cannot govern what you cannot evaluate, and we are still learning how to evaluate AI.