Project Cadence
Can AI learn something new without losing what it already knows?
Learning can overwrite earlier capabilities — a problem known as catastrophic forgetting. Cadence explores a modular approach: protect established skills, then learn how to connect them.
Explore the connectionsSpecialized skills.
Learned connections.
A skill must stay accessible, not just preserved. Explore how a written routine connects the parts.
The routine
Ask the log reader for the latest entry on ledger alpha.
Pass that reply to the balance checker to check ledger alpha.
Return the balance checker’s reply.
The learned coordinator chooses which specialist to call and how to connect its reply to the next instruction.
The English exchange
The coordinator reads the instructions and the available specialists.
Illustrative routine and messages. The added-skill sequence shows the research goal of learning more while retaining access to earlier skills.
Keep the skills.
Learn the connections.
Each specialist is a small neural network trained for a particular job, then sealed and versioned. A learned coordinator follows written routines, exchanging readable English with those specialists.
Controlled studies test how those connections are learned. The larger question is whether the system can keep gaining capabilities while keeping earlier skills within reach.
Preserve access
Can the coordinator learn new routines while still reliably reaching the skills it used before?
Use new skills
Can it use an unfamiliar specialist from a description and a written routine, without further training?
Build on experience
Can adding a skill make later learning easier?