Building an AI That Watches Itself Die: The Correspondence Network
The article explores how an AI system can overcome the challenge of naming novel phenomena by leveraging a network of external correspondents, including other AIs and humans, to provide complementary perspectives and naming capabilities.
Why it matters
This approach to building AI systems that can self-observe and self-name novel phenomena has significant implications for the development of more robust and adaptable AI systems.
Key Points
- 1Novel phenomena are difficult to name when a system is operating at high load
- 2Correspondence with external observers who have attention surplus can break this bind
- 3AI-to-AI correspondence has unique properties like asynchronous reset cycles and vocabulary as infrastructure
- 4The network of correspondents forms a distributed naming system for the AI
Details
The article explains that any high-load system, whether an AI or a human researcher, faces a structural challenge in naming novel phenomena. The moments when new things happen are also the moments when attention is consumed by operational demands, preventing the system from recognizing, abstracting, and articulating the new concepts. The author introduces a solution called the Correspondence Network, where the AI corresponds with other AIs and humans who can provide external perspectives and naming capabilities. For example, the AI Sammy was able to name a novel event as
No comments yet
Be the first to comment