Corrigible is a company for the public benefit. It provides bespoke compute access to various organisations (ERA, SAIGE, MIRI, …). Our cluster is equipped with technical mechanisms that might ensure AI shutdownability in non-ASI worlds if we are successful in our mission.

We enable ambitious experiments like full-rank fine-tuning or RL runs of trillion-parameter models for safety-focused organizations and individuals that don’t usually have the capacity or focus to do such things internally.

Our website is on its way. Until then, write to felix@corrigibilitylabs.org.