Abstract:

Ambitious projections suggest that artificial general intelligence (AGI) might emerge in the next five years. This prospect raises serious considerations about how the required leap-ahead technologies could materialize based on extending the scale of current training methods and architectures. Within the broader AI community, this AGI singularity has yielded to a more gradual transition or inflection to describe frontier intelligence or Ph.D. level generalist. The logistics of building those technologies remain uncertain. Using a technical case study approach, this paper addresses a key roadblock in corralling the necessary training data for large language models (LLMs) to grow beyond their current constraints. Although the internet-scale human textual knowledge provides over 10 trillion tokens (or words) in 2025, the quality and quantity of additional training content remains unclear. For example, tutorial and domain-specific examples have shown considerable promise to improve LLM performance but generating that scale of content requires less human assembly and more machine curation to make noticeable improvements. This approach draws inspiration from game learning via just self-play. Self-play means unmanned gaming (machine vs. machine) with verifiable outcomes (wins and losses). One approach to AGI employs an analogy to these same self-supervised learning approaches. We propose suitable problem classes for simulations that support machine-only training data and self-generated domain expertise. We explore the possibility of self-play for vast data generation strategies with a novel actor-critic game that optimizes DoD learning skills. These simulations address open-ended and complex model problems of DoD priorities including supply chain optimization, logistical resource allocation, shortest path-finding, and software code quality analysis. This approach differs from standard text augmentation with synthetic conversational data. We specialize AGI expert growth to the sixteen industries that comprise the national critical infrastructure.

Keywords: AGENT-BASED SIMULATION;AI;MACHINE LEARNING;SIMULATIONS;TRAINING

Forrest McKee

Affiliation

Refine Results

How Artificial General Intelligence Will Train Itself