How do differentiable cosmogonies relate to Microscope AI?

Similarities:

• Both avoid deploying agentic systems.
• Both rely on gaining knowledge from an ML model or its sandboxed output.
• They both likely fail to account for Meta deploying unaligned AGI six months later.
• The interpretability aspect of unfamiliar structures (e.g. ML model weights, alien replicators) appears to be the most difficult step of the proposals.

Differences:

• Differentiable cosmogonies rely on gleaning information from the simulation implemented by an ML model, while Microscope AI relies on gleaning information from the weights of the ML model.
• Microscope AI relies on training the ML model on data about our world, while differentiable cosmogonies as a paradigm rely on keeping the ML model as isolated as possible from information about our world.