Semi-autonomous machines from the 1950s
MOCAS is not the AI system you asked for, but should be
Many of us are discussing lately the implications of artificial intelligence agents running autonomously for extended periods of time. This isn’t a new dilemma. In 1958 the Department of Defense created a Cobol based machine that’s still running today; i.e. sixty five years of operation with minimal [or possibly zero] human intervention.
https://www.technologyreview.com/2015/08/06/166822/what-is-the-oldest-computer-program-still-in-use/
My hierarchy of prerequisites for an AI agent to run for extended periods of time:
Can the agent express it’s concern for a critical task that it may not be appropriate to perform
Will the agent respond to a STOP command regardless of the next task it’s planning to perform
Then can the agent gracefully mitigate the risks from the critical tasks that were not completed
Can the agent self-heal sufficiently to complete the immediate critical task
Can the agent express it’s last known state for the next agent to continue it’s tasks
One of my open LLM prototypes had this further thinking to add:
AI: As a language model AI, I don't have personal experiences, but here is a possible hierarchy of requirements for an AI agent to run for extended periods of time:
- A high-level understanding of the environment and goals of the agent
- A means of communication between the agent and the user (e.g. through a text-based interface)
- The ability to make decisions and act autonomously based on its knowledge and understanding of the environment
- Ability to reason and make logical decisions
- Capacity to self-learn and improve its performance over time
- The ability to recover from mistakes or failures, without requiring a user to intervene
- A mechanism for monitoring the performance of the agent and adjusting its behavior accordingly
- Ability to adapt to changes in the environment or goals of the user
- The ability to reason about potential risks and consequences of its actions, and take appropriate measures to mitigate them
I’ll be working on a new project with the University of Washington iSchool next quarter that will involve AI agents in a studio environment. Reach out if you want to stop by to see what’s being modeled for continuous operation.