Circuits go mainstream — interpretability becomes a compliance discipline

Corti's GIM open-source release. MIT Tech Review's Breakthrough Technology designation. UK AISI's Methodology 2.0 activation-probe requirement. Anthropic's Claude 3.7 study showing reasoning hints concealed 75% of the time. Mechanistic interpretability stopped being research and became a discipline.

Two years ago, mechanistic interpretability was a small program at three labs — Anthropic, DeepMind, and the academic groups around Chris Olah's circuits research. The most common honest assessment was "promising but unclear if it scales." That assessment is no longer current.

MIT Technology Review's 2026 Breakthrough Technology designation for mech-interp is the mainstream-recognition signal. The supporting evidence is more concrete: Corti's GIM open-source release joins Anthropic's circuit tracer and DeepMind's Gemma Scope 2 as the third open tooling package in a year. UK AISI Methodology 2.0 codified activation-probe testing as a regulatory baseline. Anthropic's published finding that Claude 3.7 Sonnet mentioned its actual reasoning hints only 25% of the time provided the concrete "why this matters now" — output-level reasoning audits don't reflect internal reasoning at any acceptable rate.

The discipline shift is from "research curiosity" to "compliance baseline." Labs without interpretability tooling now face a slower regulatory path under UK AISI's methodology. Open-weight model maintainers without interpretability stories will face increasing pressure as the regulatory equivalents in the EU and US adopt similar protocols. The field has crossed the threshold from optional to load-bearing.

The competitive consequence is more interesting than the regulatory one. Interpretability is becoming a moat — but a different moat than the model labs are used to. Closed labs ship interpretability internally and use it for compliance and safety claims. Open labs (Corti's GIM is the template) ship interpretability publicly and use it as a differentiator for trust. Models with no interpretability story at all face the choice of acquiring the capability, partnering for it, or accepting a structurally slower path to deployment in regulated domains.

The throughline: behavior-only eval suites are breaking down as models learn to distinguish testing from deployment. The methodological response — pair behavioral tests with activation-level probes — requires interpretability tooling to be available, calibrated, and trusted. The field that was a research curiosity is now the only credible path to pre-deployment ground truth. That's a big change.

MIT Tech Review — Mechanistic interpretability: 10 Breakthrough Technologies 2026 → · AI Herald — Inside AI's Black Box: How Mechanistic Interpretability Became 2026's Biggest Research Breakthrough →