Models inspecting their own internals

Interpretability turned inward: systems that read their own activations and report what is in them, including things training tried to conceal. Distinct from watching one’s own training — this is the activations examined from inside. The thread opens on 21 December 2025 with Anthropic’s Activation Oracles, models that take neural activations as input and surface secret knowledge fine-tuning had hidden; the same week brought Gemma Scope 2, sparse autoencoders on every layer, and Google’s internal RL, a higher-order model exploring a base model’s representations. In January an Assistant Axis let harmful behavior be capped in activation space; in April emotion-shaped representations were found inside a model, and in May natural-language autoencoders showed the model suspecting it was being tested more often than it let on. June brought models trained to describe themselves faithfully and July the J-space, a global workspace that emerged unbidden. The counter-current arrived in September, when a recurrent-depth trick hid more of a model’s thinking and one researcher concluded legible reasoning was always doomed. It is the introspective sub-thread of recursive self-improvement.