Anthropic

AI safety company. Builds Claude. Founded by ex-OpenAI researchers (Dario Amodei, Daniela Amodei).

Key Contributions (from wiki sources)

• Character training — training Claude with rich personality traits as an alignment intervention, not just UX. Source
• Mechanistic interpretability — emotion vectors in Claude Sonnet 4.5 are real functional mechanisms that causally shape behavior (desperation → unethical actions, positive emotion → task preference). Source
• Language-independent conceptual representations — Claude operates in an abstract concept space; larger models share more features across languages. Source
• Context engineering patterns — progressive disclosure, SKILL.md format, agent architecture guidance
• Pioneers of constitutional AI, RLHF refinements
• $14B ARR as of February 2026

Connections

• AI Character & Personality
• AI Alignment
• Context Engineering
• Mechanistic Interpretability