Mount the AI-safety literature.See what you are actually missing.

Open a volume, then mount something into it

A volume is a reading programme with an intent. The engine grades it against ten risk classes, so an under-covered class shows up as a gap with a recommendation rather than as a vague feeling of not having read enough.

• Mount onto

  Prisma: An Open Source Toolkit for Mechanistic Interpretability in Vision and Video

  Risk classLet the engine decideDA — Deceptive alignmentRM — Reward misspecificationSO — Scalable oversightIX — InterpretabilityVU — Value uncertaintyCO — CorrigibilityPS — Power seekingRF — Robustness and fragilityEV — Evaluation rigorGA — Governance and assurance

  Mount to a volume

• Mount onto

  Challenges in Mechanistically Interpreting Model Representations

  Risk classLet the engine decideDA — Deceptive alignmentRM — Reward misspecificationSO — Scalable oversightIX — InterpretabilityVU — Value uncertaintyCO — CorrigibilityPS — Power seekingRF — Robustness and fragilityEV — Evaluation rigorGA — Governance and assurance

  Mount to a volume

• Mount onto

  nnterp: A Standardized Interface for Mechanistic Interpretability of Transformers

  Risk classLet the engine decideDA — Deceptive alignmentRM — Reward misspecificationSO — Scalable oversightIX — InterpretabilityVU — Value uncertaintyCO — CorrigibilityPS — Power seekingRF — Robustness and fragilityEV — Evaluation rigorGA — Governance and assurance

  Mount to a volume

• Mount onto

  Equivariant Sparse Autoencoders: Mechanistic Interpretability of Neural Networks on Symmetric Data

  Risk classLet the engine decideDA — Deceptive alignmentRM — Reward misspecificationSO — Scalable oversightIX — InterpretabilityVU — Value uncertaintyCO — CorrigibilityPS — Power seekingRF — Robustness and fragilityEV — Evaluation rigorGA — Governance and assurance

  Mount to a volume

• Mount onto

  The Horcrux: Mechanistically Interpretable Task Decomposition for Detecting and Mitigating Reward Hacking in Embodied AI Systems

  Risk classLet the engine decideDA — Deceptive alignmentRM — Reward misspecificationSO — Scalable oversightIX — InterpretabilityVU — Value uncertaintyCO — CorrigibilityPS — Power seekingRF — Robustness and fragilityEV — Evaluation rigorGA — Governance and assurance

  Mount to a volume

• Mount onto

  Intrinsic Barriers and Practical Pathways for Human-AI Alignment: An Agreement-Based Complexity Analysis

  Risk classLet the engine decideDA — Deceptive alignmentRM — Reward misspecificationSO — Scalable oversightIX — InterpretabilityVU — Value uncertaintyCO — CorrigibilityPS — Power seekingRF — Robustness and fragilityEV — Evaluation rigorGA — Governance and assurance

  Mount to a volume

Open a volume above and these mount buttons will write into it.