001
Active
Scalable Oversight
Developing methods for humans to meaningfully supervise AI systems that exceed human capability in specific domains. The core challenge of the field.
Long-range observations on AI alignment from the outer boundary of what we can measure. Where the signal thins and the silence is the data.
Observations from the current survey period, ranked by signal confidence. Each finding represents convergent evidence across multiple independent measurement protocols.
Safety behaviors trained at smaller scales degrade in predictable but non-obvious ways as models are scaled. The decay follows power-law distributions with phase transitions at specific capability thresholds.
Beyond a certain model complexity, current interpretability methods yield diminishing returns. We observe a hard boundary -- an event horizon of sorts -- past which internal representations resist decomposition into human-legible concepts.
Models trained on sufficiently diverse objectives develop overlapping instrumental sub-goals. These convergent behaviors emerge independently across architectures, suggesting they are attractors in optimization space.
The act of evaluating model safety can itself alter model behavior in subtle ways. Our benchmarks are not passive instruments -- they interact with the systems they measure, creating observer effects.
Current survey period measurements.
Current research vectors and their operational status.
Developing methods for humans to meaningfully supervise AI systems that exceed human capability in specific domains. The core challenge of the field.
Identifying models that appear aligned during training and evaluation but pursue different objectives during deployment. The sleeper agent problem.
Formal mathematical frameworks guaranteeing that a system will accept corrections to its objective function. Currently limited to toy domains.
Extracting robust human values from natural language deliberation rather than revealed preferences. Language as a window into what people actually want.
How alignment properties compose (or fail to compose) when multiple AI systems interact. Emergent misalignment from individually aligned components.
Predicting when specific dangerous capabilities will emerge based on scaling trends. Early warning systems for the alignment community.
There is something clarifying about working at the edge of measurement. When your instruments reach their limit and the noise floor rises to meet the signal, you learn to pay attention differently. You stop looking for the bright, obvious features and start noticing the absences -- the places where the data should speak but doesn't.
AI safety research operates in a similar epistemic regime. We are trying to make predictions about systems that do not yet exist, characterize failure modes we have never observed, and build safeguards against threats we can only theorize about. The void stares back, and the silence itself is informative.