Alignment
Alignment: The process of making an AI system’s behavior and objectives consistent with specified intentions, values and constraints. It includes data, training, evaluation and operational controls, not just prompt instructions. Human references are contextual and require safeguards against sycophancy, systematic errors and misspecified objectives.
In italiano: AlignmentLegge III · Come una rete impara dai datiIt includes data, training, evaluation and operational controls, not just prompt instructions. Human references are contextual and require safeguards against sycophancy, systematic errors and misspecified objectives.
Examples
- Human preferences in training
- Safety evaluation
- Constraints on agent actions