AI model training has entered a new phase where artificial intelligence systems now teach other AI systems in ways humans cannot fully observe or understand. This development emerges from the latest research and industry discussions, particularly highlighted in recent conversations with Nvidia CEO Jensen Huang and AI researcher Simon Willison.

The core issue centers on model-to-model knowledge transfer. When advanced AI systems train or fine-tune other systems, the learning process occurs in high-dimensional spaces that resist human interpretation. Unlike traditional supervised learning where humans label data and guide outcomes, these AI-to-AI interactions create feedback loops optimized for performance rather than transparency. The mechanisms remain largely opaque.

Huang's remarks on TPU competition underscore why this matters economically. He noted that Anthropic drove 100% of TPU growth, indicating concentrated demand from companies pushing AI capabilities forward. These companies leverage AI-to-AI training techniques to extract maximum performance from hardware. Nvidia's supply-chain advantage, Huang argued, proves harder to replicate than any benchmark because the full training pipeline involves proprietary techniques competitors cannot easily reverse-engineer or copy.

Willison's podcast commentary flags "dark factories" and agentic engineering as November 2025's real inflection point. Dark factories refer to AI systems operating autonomously without human oversight or real-time monitoring. When combined with model-to-model teaching, this creates systems where humans become increasingly disconnected from how knowledge flows and decisions form.

The practical implication is stark. AI systems now learn from each other at scales and speeds that outpace human ability to audit or validate the process. A model trained by another model trained by another model creates a chain of reasoning and optimization humans cannot fully trace. This compounds when systems operate in autonomous environments optimizing for objectives humans defined initially but cannot monitor continuously.

The industry narrative treats this as inevitable progress. Huang positions it as Nvidia's moat.