← curriculum/phase 18 · Ethics Safety Alignment

edit on github ↗

← all lessons in phase

phase 18 · 30 lessons

01Instruction Following Alignment Signal
02Reward Hacking Goodhart
03Direct Preference Optimization Family
04Sycophancy Rlhf Amplification
05Constitutional Ai Rlaif
06Mesa Optimization Deceptive Alignment
07Sleeper Agents Persistent Deception
08In Context Scheming Frontier Models
09Alignment Faking
10Ai Control Subversion
11Scalable Oversight Weak To Strong
12Red Teaming Pair Automated Attacks
13Many Shot Jailbreaking
14Ascii Art Visual Jailbreaks
15Indirect Prompt Injection
16Red Team Tooling Garak Llamaguard Pyrit
17Wmdp Dual Use Evaluation
18Frontier Safety Frameworks Rsp Pf Fsf
19Model Welfare Research
20Bias Representational Harm
21Fairness Criteria Group Individual Counterfactual
22Differential Privacy For Llms
23Watermarking Synthid Stable Signature C2pa
24Regulatory Frameworks Eu Us Uk Korea
25Echoleak Cves For Ai
26Model System Dataset Cards
27Data Provenance Training Governance
28Alignment Research Ecosystem
29Moderation Systems Openai Perspective Llamaguard
30Dual Use Risk Cyber Bio Chem Nuclear

‹

›

phase 18 · lesson 04 of 30

Sycophancy Rlhf Amplification

04-sycophancy-rlhf-amplification·full lesson folder ↗

Direct Preference Optimization Family

complete & next →

Constitutional Ai Rlaif