Akili – gated skill updates and rollback receipts for deployed models (github.com)

Akili has announced a significant advancement with the release of v0.1.0-rc4, introducing a versioned, continual-capability runtime designed for safe and efficient adaptation of deployed AI models. This platform enables models to learn from new experiences while preserving safety and resource constraints by isolating learned capabilities and validating them before activation. The system manages the lifecycle of capabilities, including active, dormant, and rolled-back states, ensuring a tamper-evident path that records each certified safe state. This capability enhances operational reliability and reliability in production environments. The Akili runtime is notable for its pluggable learning engines, which allow for flexibility in adapting to new challenges without relying on a single, monolithic algorithm. Key technical details include rigorous benchmarking results across various metrics, such as achieving a 99.5% overall accuracy in skill evaluations during experiments. The platform's architecture supports ongoing learning while enabling secure rollback to previous states if new capabilities prove unsafe. This innovation positions Akili as a valuable tool for the AI/ML community, particularly in fields where safety and adaptability are critical.