Skip to content
Sphere · Whitepaper

AI Alignment

The Alignment Problem in AI development refers to the challenge of ensuring that artificial intelligence systems act in accordance with human values and intentions, even as they become more advanced and autonomous.

The Centralization Risk

Section titled “The Centralization Risk”

A critical issue is the problem of centralized powers that could monopolize access to powerful superhuman AI and make it serve their interests instead of considering every human and every life.

Current alignment techniques (training on human text, fine-tuning with feedback, reward modeling) will not be sufficient at scales approaching pseudo-sentience or real consciousness.

Decentralized Alignment

Section titled “Decentralized Alignment”

In order for AI to properly represent human will, it is important to decentralize the power of the technology on two levels:

1. Access to Technology

Section titled “1. Access to Technology”

Open-source projects and datasets are especially important, as they enable access to anyone with the proper skills and resources. See Open-Source Development.

2. Democratic Representation

Section titled “2. Democratic Representation”

Since not all people can use or develop AI tools directly, it is important to create democratic systems that efficiently abstract the opinions of large groups into a concise representation of the group’s will.

The Sphere Approach

Section titled “The Sphere Approach”

The Trust-Based Social Network could provide a basis for AI alignment protocols:

  • Governance Engines enable complex and dynamic voting and polling that feeds into AI training
  • Value Networks define value objects with legal descriptions and training data
  • Decentral Voting Systems ensure that alignment reflects genuine democratic consensus
  • Eventually, cryptographic networks could distribute the operation of AI services to a network of individuals, instead of centralized servers

The Urgency

Section titled “The Urgency”

It is of utmost importance to unify our values and opinions among ourselves before constructing superhuman AI, as conflicts between people with access to these powerful entities could lead to catastrophic and unpredictable consequences.

Section titled “Related Concepts”