Reading Lists par Niveau

Débutant (0-20h)

Objectif: Comprendre qu'il y a un problème

1. AGI Ruin: A List of Lethalities (Eliezer)

• https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a-list-of-lethalities
• 2h lecture

2. AI Alignment: Why It's Hard, and Where to Start (Video - Eliezer)

• https://www.youtube.com/watch?v=EUjc1WuyPT8
• 1h

3. Rob Miles YouTube Channel (vulgarisation)

• https://www.youtube.com/@RobertMilesAI
• 5-10 videos (5h)

4. The Alignment Problem (Brian Christian - livre)

• Accessible, journalistique
• 10-15h

---

Intermédiaire (20-100h)

Objectif: Comprendre problèmes techniques principaux

1. Risks from Learned Optimization (Hubinger et al.)

• https://arxiv.org/abs/1906.01820
• 10h (paper + discussions)

2. Concrete Problems in AI Safety (Amodei et al.)

• https://arxiv.org/abs/1606.06565
• 5h

3. Embedded Agency (Sequence)

• https://www.alignmentforum.org/s/Rm6oQRJJmhGCcLvxh
• 20h

4. Superintelligence (Nick Bostrom - livre)

• Classique, un peu daté (2014) mais fondamental
• 20h

5. ELK Document (ARC)

• https://docs.google.com/document/d/1WwsnJQstPq91_Yh-Ch2XRL8H_EpsnjrC1dwZXR37PC8
• 10h

6. Alignment Forum (curated posts)

• https://www.alignmentforum.org/
• 20-30h (sélection)

---

Avancé (100-500h)

Objectif: Comprendre research frontiers, contribuer

1. MIRI Research

• https://intelligence.org/research/
• Agent foundations papers
• 50-100h

2. Corrigibility (MIRI)

• https://intelligence.org/files/Corrigibility.pdf
• 20h (+ related work)

3. Logical Induction (MIRI)

• https://intelligence.org/files/LogicalInduction.pdf
• 40h (très math-heavy)

4. Iterated Amplification (Paul Christiano, all posts)

• https://ai-alignment.com/
• 30h

5. Constitutional AI + Mechanistic Interpretability (Anthropic research)

• https://www.anthropic.com/research
• 40h

6. Debate, ELK, related (ARC, OpenAI)

• Papers + discussions
• 50h

7. Alignment Forum (comprehensive reading)

• Major sequences, debates
• 100-200h