Interpretability
Opening the black box: what a model represents, where it stores it, and how that representation changes under pressure.
Read the programme →Frontier research
The responsible development of transformative AI systems.
We study the fundamental nature of intelligence to build AI that is interpretable, steerable, and aligned with human values. Our work is guided by a single conviction: the most capable systems must also be the most understood.
Read our research →Research areas
03 active programmesSteerability
Building systems that can be corrected at the level of intent, not just output — and can explain why they changed course.
Read the programme →Alignment
Measuring whether a system's behaviour remains within human values when the context shifts beyond its training distribution.
Read the programme →Meridian Research · Frontier AI← 模板展厅