Aligning models with human intent
We build systems that are helpful, honest, and harmless. Our alignment research focuses on ensuring our models understand and follow complex human instructions safely.
Our alignment pillars
The core methodologies we use to train and evaluate our models.
Reinforcement Learning from Human Feedback (RLHF)
We use RLHF to fine-tune our models based on human preferences. By training reward models on thousands of human comparisons, we teach our systems to prioritize responses that are helpful and accurate.
Constitutional AI
We provide our models with a set of principles or a "constitution" to guide their behavior. This allows the model to self-critique and revise its own responses to ensure they align with safety guidelines.
Red Teaming
We employ dedicated teams of experts to actively try to break our models. This adversarial testing helps us discover edge cases, biases, and vulnerabilities before they reach production.
Interpretability
We are investing heavily in mechanistic interpretability—understanding the internal representations of our models. By looking inside the "black box," we can better predict and control model behavior.