wtf is akili-runtime?

Akili Runtime is a research project exploring how an AI system could learn new skills throughout its life without either overwriting what it already knows or being tricked by bad updates, an area researchers call continual learning. The core idea is a control loop: new experience becomes a candidate skill, that candidate goes through a safety check before being allowed into active use, the system tracks each skill as active, dormant, or rolled back, and every change is written to a tamper evident log so any of the system's past safe states can be recovered. The project describes itself as a runtime that can plug in different learning methods rather than one single learning algorithm, and states plainly that it is not claiming to have invented adapters, replay buffers, model registries, or rollback mechanisms individually, only that it combines them into one tested system. Two main experiments back its claims so far. A small simulated robotics task run on ordinary CPUs across three repeated trials showed the system successfully learning four skills per trial, correctly rejecting one unsafe update while keeping the previous safe version running, and completing certified multi step missions. A second, GPU based experiment simulating agents that evolve over time reported that Akili caught and blocked a deliberately planted malicious update while a comparison system without those safeguards did not. The project also ran a matched comparison against an existing continual learning method called DER++ on a standard image classification benchmark, reporting notably less forgetting of earlier learned material, though the authors note that a simpler baseline came close on that same test, and describe the current results as evidence still working toward a stable, fully released version. The code is written in Python, installs through a standard pip virtual environment, and includes a short example script along with a full automated test suite. The repository is currently tagged as a release candidate rather than a finished first version, with additional benchmark data still pending import.