Steering group sets four initial thrusts
On October 7, ICINAi’s founding steering group met and identified four initial thrusts: shared resources that interpretability research needs and that no single lab can build alone.
Frontier model access
A trusted procedure for running experiments on closed frontier models, and hosting for the largest open-weight models through NDIF.
What it’s forStudy the models people actually use.
Benchmarks and model organisms
Shared benchmarks and models with known hidden behaviors, starting from the model organisms of the Aletheia’s Quest lie-detection competition.
What it’s forTest which methods are effective at monitoring latent intents.
An atlas of mechanisms
A shared registry of mechanisms found inside neural networks, each indexed by the test data and causal experiments that isolate it.
What it’s forConnect results across labs and models.
Domain experts
A program pairing leading scientists with research projects on the gap between what AI knows and what people know.
What it’s forTurn superhuman AI knowledge into human insight.
Working groups for each thrust are forming now. We are seeking funding and collaborators to support these efforts.
Get involved