Concepts

Short pages for terms that come up in several lectures. Each one has a definition, the key idea or formula and an Appears in list that links to the detailed explanation in the lectures. The list below grows with every lecture.

LLM Foundations

• Transformer
• Tokenization
• Mixture of Experts
• Linear Probe
• Decoding and Sampling
• Training Compute
• Supervised Fine-Tuning
• RLVR
• LoRA
• Representation Engineering
• Assistant Persona

Attacks

• Adversarial Example
• Adaptive Attack
• Jailbreak
• Prompt Injection
• GCG
• PAIR
• Distillation Attack
• Refusal Direction
• Data Poisoning

Defenses

• Adversarial Training
• Swiss-Cheese Model
• Unlearning

Transparency and Privacy

• Watermarking
• AI Text Detection
• Content Provenance
• Memorization
• Privacy Threat Models
• Contextual Integrity

Alignment

• AI Alignment
• RLHF
• DPO
• Constitutional AI
• LLM as a Judge
• Deliberative Alignment
• Safe-Completions
• Reward Hacking
• Emergent Misalignment

Agents

• LLM Agent
• Model Context Protocol
• Agent Skills
• Lethal Trifecta
• Agent Design Patterns

Multi-Agent Safety

• Multi-Agent Risks
• Infectious Jailbreak
• Algorithmic Collusion
• Steganography
• Sycophancy

Deception and Control

• Scheming
• Evaluation Awareness
• Alignment Faking
• Sandbagging
• Chain-of-Thought Monitoring
• AI Control
• Scalable Oversight

AI R&D Automation

• Recursive Self-Improvement
• AI R&D Benchmarks

Forecasting

• METR Time Horizon
• Scaling Laws
• Forecasting and Calibration