Alignment RL Team
The Alignment RL team focuses on developing and implementing new alignment techniques for language models, with the goal of improving model values, honesty, and character.
- Implements and scales techniques like oversight, synthetic data generation, and training models to assist in model training
- Works at the intersection of alignment research and production systems
- Collaborates with other teams like Safeguards and Alignment Science
- Evaluates the effectiveness of alignment interventions
- Develops prompting strategies and data generation pipelines to improve model responses
- Identifies and fixes edge case behaviors through rigorous testing
The team includes prompt engineers and policy designers who shape AI system behavior to ensure alignment with human values through prompt engineering, ethical judgment, and knowledge of diverse scenarios.
Production Model Post-Training Team
The Production Model Post-Training team is responsible for turning our base models into the production Claude models that millions of users interact with daily.
- Directly trains the flagship models launched to the public via Claude.AI and Anthropic's API
- Designs and iterates on state-of-the-art finetuning techniques, such as Constitutional AI and RLHF
- Implements new algorithms, runs experiments on data mixes, and designs evaluations
- Improves production model training pipelines
- Translates novel finetuning techniques into production model training processes
- Ensures models are helpful, honest, and harmless
- Develops robust evaluations for new model capabilities
The team directly impacts the Claude models that millions of users experience daily, combining state-of-the-art research with production-ready implementations.
Horizons Team
The Horizons team leads Anthropic's reinforcement learning research and development, playing a critical role in advancing AI systems' capabilities while ensuring safety.
- Develops systems that enable models to use computers effectively
- Advances code generation capabilities through reinforcement learning
- Pioneers fundamental RL research for large language models
- Builds scalable RL infrastructure and training methodologies
- Enhances model reasoning capabilities in areas like mathematics and programming
- Collaborates closely with alignment and frontier red teams to ensure systems are both capable and safe
- Partners with the applied production training team to bring research innovations into deployed models
- Works with dedicated RL engineering teams to implement research at scale
- Has contributed to all Claude models, with significant impacts on the autonomy and coding capabilities of more recent models
The Horizons team sits at the intersection of cutting-edge research and engineering excellence, with a deep commitment to building high-quality, scalable systems that push the boundaries of what AI can accomplish.
Knowledge Team
The Knowledge team works on improving how language models interact with external information.
- Performs finetuning and reinforcement learning to teach language models how to interact with new information architectures
- Builds "hard" knowledge base evaluation sets to identify failure modes in how language models work with external data
- Designs and evaluates advanced agentic search capabilities
- Develops scalable distributed information retrieval systems, including:
- Search engines
- Knowledge graphs
- Retrieval-Augmented Generation (RAG)
- Indexing and ranking systems
- Query understanding
- Distributed data processing
The Knowledge team bridges the gap between language models and external information sources, enabling more accurate and up-to-date AI interactions.
Agents Team
The Agents team focuses on developing reliable AI agents and frameworks for building trustworthy AI systems.
- Researches agent capabilities and safety considerations
- Develops technical frameworks for building trustworthy AI
- Creates simple, composable patterns for implementing LLM agents across industries
- Designs tools that enable Claude to interact with external services and APIs
- Develops agent-computer interfaces (ACI) with clear documentation and examples
- Tests and iterates on how models use tools for various tasks
- Implements verification systems to ensure agent outputs meet quality standards
- Explores both fully autonomous systems and more prescriptive implementations that follow predefined workflows
The Agents team focuses on making AI systems more capable and reliable while ensuring they remain trustworthy and aligned with human intentions.