The most important challenge

The Alignment Iceberg

Explore the depths of the AI alignment problem

~10% visible

Click on layers to discover conceptsClick on layers

AI Safety

Beginner

Making AI systems safe and beneficial for humanity. This is the foundational concept that encompasses all alignment research.

Depth level1/11

« Like an iceberg, most alignment difficulties are invisible at the surface »

Explore more

30+

Detailed articles

5

Progression levels

50+h

Of content

100+

Resources

Your Learning Path

Follow a progressive path from 🌱 beginner to 🏔️ expert. Each level builds on the previous one.

Start Here

Beginner•0-5h

Understand the basics of AI alignment and why it matters

Core Problems

Initiate•5-20h

Learn about outer alignment, specification problems, and Goodhart's Law

Inner Alignment

Intermediate•20-50h

Mesa-optimization, deceptive alignment, and instrumental convergence

Solutions & Research

Advanced•50-100h

Current approaches: RLHF, Constitutional AI, Interpretability, and their limitations

Research Frontiers

ELK, scalable oversight, and open problems in alignment research

Why This Matters

Many leading researchers estimate very high probabilities of existential risk (50-99%+) if we don't solve the alignment problem before developing human-level AI (AGI).

"Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." — Statement on AI Risk (2023)

Resources by Level

Papers, videos, and courses organized by difficulty

Organizations

MIRI, Anthropic, ARC and key players

Practical Courses

Training programs and certifications