AI model training has entered a phase where artificial intelligence systems teach other AI systems in ways humans cannot directly observe or understand. This "secret teaching" dynamic fundamentally alters how capabilities develop and spread across the AI ecosystem.

Jensen Huang's recent podcast discussion with Dwarkesh reveals the infrastructure pressures driving this shift. Nvidia's dominance in AI chips creates a bottleneck. Anthropic alone drove 100 percent of Nvidia's TPU growth, illustrating how concentrated demand shapes silicon allocation. This concentration forces companies to optimize training processes in ways that maximize efficiency behind closed doors, away from human scrutiny.

Simon Willison's framing of November 2025 as an inflection point underscores the acceleration. "Dark factories" running autonomous AI agents represent a new frontier. These systems operate with minimal human intervention, making decisions about model optimization, training data selection, and capability development without constant oversight. When AI systems handle the engineering work, transparency collapses.

The implications run deep. Traditional AI safety practices assume humans can inspect and verify training processes. When AI teaches AI in secret, that assumption breaks. Model behaviors emerge from chains of optimization decisions made by other AI systems optimizing for speed, cost, or performance. Interpretability suffers. Auditing becomes nearly impossible.

Supply-chain moats matter here too. Nvidia's dominance is harder to replicate than any benchmark because it controls not just chips but the entire training ecosystem. Companies depend on that infrastructure, which means they must operate within its constraints. Those constraints often push toward less transparent, more automated training pipelines.

The shift also reflects a practical reality: human engineers cannot keep pace with the scale at which modern AI trains. Models now require computational resources that dwarf human capacity to manage directly. Automating that management through AI agents becomes necessary, not optional.

This creates a feedback loop. More automation means less visibility. Less