Corti releases GIM — open-source benchmark-leading mechanistic interpretability method for circuit discovery
Corti released GIM on May 6, an open-source mechanistic interpretability method that the team claims sets a new benchmark on AI circuit discovery for transformer models. The release lands at a moment when MIT Technology Review formally recognized mech-interp as a 2026 Breakthrough Technology and when UK AISI's Evaluation Methodology 2.0 made activation-probe protocols a regulatory baseline.
GIM joins Anthropic's open-source circuit tracer and Google DeepMind's Gemma Scope 2 as the third major open release in a year for mech-interp tooling. The field has shifted from "a research curiosity at three labs" to "a public-tooling competition with regulatory implications" faster than most expected. Corti's release means open-weight model maintainers now have a path to comply with the AISI Methodology 2.0 activation-probe requirement without building tooling in-house.
The strategic consequence is that interpretability is becoming a moat — but a different kind of moat than the model labs had previously built. Closed labs can ship interpretability internally and use it for compliance; open labs can ship interpretability publicly and use it as a differentiator. Models with no interpretability story at all are now in the position of having to either acquire the capability or accept a slower regulatory path.
Corti — GIM: A new standard for mechanistic interpretability → · 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 →