Safe, Productive, Efficient, Accurate, and Responsible (SPEAR) AI Systems
This is a special edition of my Substack AI newsletter. We just uploaded my working paper on Safe, Productive, Efficient, Accurate, and Responsible (SPEAR) AI systems, so I wanted to provide the direct link. I conducted a thorough review of published research on LLMs, which represents the first 10 pages, followed by a description of our KOS, which serves as a good example of a SPEAR AI system.
https://kyield.com/images/SPEAR_AI.pdf
Safe, Productive, Efficient, Accurate, and Responsible (SPEAR) AI Systems
Mark Montgomery
KYield, Inc.
Working Paper Copyright © 2023/2024 by Mark Montgomery
Abstract
Since the first large language model (LLM) chatbot was released to the public, leading experts in AI, catastrophic risk, economics, and cybersecurity, among others, have warned about the unprecedented risks caused by interactive LLM bots trained on large-scale, unstructured data [[i], [ii]]. Although theoretical methods have been proposed, and incremental improvements are occurring, to date none have proven as effective or comparable to other safety-critical industries such as biological contagions or nuclear power [[iii], 5]. We therefore have an urgent need to adopt AI systems based on the proven laws of physics and economics without sacrificing the many benefits of AI. This paper is focused on the inefficiencies, risks, and limitations of LLMs, the business and economic incentives influencing decisions, and the architecture required to provide safe, productive, efficient, accurate, and responsible (SPEAR) AI systems. One such system is described—our data and human-centric KOS (EAI OS).
KEYWORDS: Artificial Intelligence, Data Governance, Risk Management, Organizational Management, Systems Engineering, Safety, Sustainability, Disaster Management, Catastrophes, Existential Risk, Bioweapons, Large Language Models, Cybersecurity