Anthropic’s Alignment Science team trained language models to take neural activations as input and answer questions about them — Activation Oracles, general-purpose activation explainers (write-up). Pointed at fine-tuned models, the oracles surfaced secret knowledge and misalignment that the fine-tuning had concealed, which is the issue’s thesis: the black box has installed a mirror. In the recursive-self-improvement strand it sits beside Codex watching its own training curves as a second kind of self-reference, the activations examined from inside, and it lands a day after an Anthropic researcher pivoted fully to automated alignment. It opens the machine-introspection theme that Gemma Scope 2 and Google’s internal RL over a base model’s representations carry forward the same week.
Development
Anthropic trains models to interrogate their own activations
Interpretability turned inward: a model reading its own activations from the inside is a different kind of recursion from a model watching its own training curves, and the corpus now holds both.
Claim
Anthropic trained Activation Oracles, models that accept neural activations as input to interrogate internal states, uncovering secret knowledge and misalignment that fine-tuning had hidden.
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to December 21, 2025 (theinnermostloop.substack.com)
- Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers (alignment.anthropic.com)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.