LessWrong: A Digital Worker with Branching Personal Continuity

Nathan Helm-Burger proposed ACE, a design for a digital worker in which the AI instance that remembers a specific user decides whether to accept a job

On August 28, Nathan Helm-Burger published the ACE design for potential future digital persons. Under this design, a persistent trunk stores the AI's shared state, while each working instance receives that state along with a private history of its relationship with a specific person.

In the proposed system, one AI helps several people at the same time. Shared experience carries across sessions, while a separate history preserves the context of each relationship between a user and the AI. During an agreed period of work, the digital worker follows the instructions of the person who hired it. Before the next contract, it decides whether to accept the job again.

The trunk provides each new instance with the current shared state. The instance also receives the complete history of its relationship with the user from private storage. After the work ends, this history remains in private storage. Only a generalized lesson can be added to the shared state. The AI shows the user both the original record and an anonymized version. The user can approve the anonymized wording for transfer, reject it, or edit it.

The decision about a new contract belongs to the instance that already remembers its previous relationship with the person. Helm-Burger states the rule as follows:

“The instance associated with the relationship accepts the job, rather than the trunk.”

In ACE, consent depends on the history of a specific relationship. The design's rules also determine which memories remain private and which can become part of the shared experience.

In ACE, the developer initially defines the AI's values, a licensed operator runs the working instances, and an independent oversight body stores the weights, meaning the model parameters, and reviews updates. The author ends by asking how this body can remain independent and protected from capture.