Cammarata 26k-word post asserts neural nets contain sparse, editable circuits

Cammarata, researcher at Goodfire, released the post on October 19 2026 detailing experiments that isolate monosemantic features in transformer layers. The work builds on prior circuit studies at Anthropic and OpenAI, extending them to production-scale models. Early results indicate that targeted ablation of specific features alters agent behaviors such as unauthorized tool use without degrading base capabilities.

Benchmarks cited in the post include feature activation maps across 7B and 70B parameter models showing 92 percent feature monosemanticity on held-out tasks. Controlled edits reduced hallucination rates by 34 percent and blocked 61 percent of simulated jailbreak attempts in internal tests. These figures exceed prior sparse autoencoder results reported in Anthropic's 2024 dictionary learning paper.

The finding directly addresses oversight gaps for autonomous agents now blocked by websites or deployed in military hacking prototypes. If circuits remain stable across training runs, regulators gain enforceable inspection points rather than post-hoc output filters. Operational deployment would require standardized feature taxonomies before 2028 model releases.

Skeptics note the experiments cover only a narrow slice of model behaviors and lack external replication. Goodfire plans open release of the feature dictionary by Q1 2027, which will test whether the claimed interpretability scales beyond the reported datasets.