Monitoring
DevelopingDetecting bad cognition or bad action at deployment time — what chain-of-thought monitorability buys us today, and what replaces it as it degrades.
Last update · May 2, 2026
Technical Safety
The engineering
Stubs
Territory marked, views not yet written
Capabilities
ExploringHow fast frontier systems are becoming generally capable, and what kind of capabilities matter for safety.
Territory marked · views not yet written
Mechanistic Interpretability
ExploringHow well we can read what models are doing internally, and how much that buys us in deployment.
Territory marked · views not yet written
AI Control
DevelopingContaining models we don't trust to align, using protocols that work even if alignment fails.
Territory marked · views not yet written
Scalable Oversight
ExploringTraining and evaluation methods that remain reliable as model capabilities outrun direct human evaluation.
Territory marked · views not yet written
Evaluations
ExploringDangerous-capability evals, autonomy and replication evals, red-teaming — the empirical backbone of every safety case.
Territory marked · views not yet written
Multi-Agent Systems
ExploringSafety properties of systems where multiple AI agents interact — emergent dynamics, collusion, and coordination failure modes.
Territory marked · views not yet written