<aside> 💡 The specification for an AI conceptual engineer

https://github.com/jerhadf/AI-CE

</aside>

Overview

Purpose

The AI Conceptual Engineer aims to assist and automate the process of conceptual engineering. This system would be a “copilot for the mind,” to automate and assist with the CE process. Using the vast linguistic training data of LLMs, along with more human input, it could build a deep and broad understanding of the concept. It could take a concept like <honesty> as input, and then decompose a concept, figure out its current function in our conceptual system, find its defects and benefits, and surface tensions or inconsistencies. Then it could help you improve the concept, explore alternatives, and generate new concepts. It could also help build a measurement system for the concept, gather data, run simulations and experiments, and elicit human feedback to empirically test and refine conceptual proposals. The goal is to augment and scale human conceptual engineering, in order to accelerate progress on conceptual issues related to AI alignment, fairness, transparency, and more.

Scope

The scope of this document includes the product outline, features, MVP stages, technical implementation, and a draft roadmap.

Features

Input and Initial Processing

• Concept Input: Accepts a single concept (e.g., "honesty") as input.
• Concept Decomposition: Breaks down the input concept into its constituents - its sub-parts or conceptual primitives.
• Function Identification: Identifies the current functional role of the concept in the relevant conceptual or social system.
• Defect and Benefit Analysis: Points out the defects and benefits of the concept.

Output: Applying Operations in CE

• Improvement Strategies: Suggests methods to improve the concept.
• Alternative Concepts: Generates and explores alternatives.
• Measurement System: Proposes a way to measure the concept's effectiveness.
• Gradual stiffening: weaving a structure which starts out globally complete, but flimsy; then gradually making it stiffer but still rather flimsy, then finally making it generally strong
• Abolition: removal of the concept from thought and talk
• Conceptual refactoring: changing which concepts are used in major social systems. Once the misfit concept has been identified, we backtrack to find where it is used or referenced in our thought, talk, and action, and replace it with an alternative, better pattern