Origins
Anthropic was founded in 2021 by Dario Amodei (former VP Research at OpenAI), Daniela Amodei, and colleagues. Core concern: as AI becomes more capable, ensuring it remains safe and aligned with human values is the critical unsolved problem.
Constitutional AI
Constitutional AI (CAI) gives the model explicit principles and asks it to critique and revise its own responses against them. Result: more self-consistently helpful and harmless behaviour with less dependence on human labeller scale.
The Claude model family
Claude 3 (2024) in three tiers: Haiku (fastest), Sonnet (balanced), Opus (most capable). Claude 3.5 Sonnet outperformed GPT-4 on several benchmarks. Notable for 200k token context windows and nuanced instruction-following.
Interpretability research
Anthropic publishes important mechanistic interpretability work — understanding what happens inside neural networks at the circuit level. Their superposition papers are landmark contributions to understanding how concepts are encoded in LLMs.
