Mechanistic Interpretability, Frontier Scale
Reverse-engineering the internals of production models.
Tier Grades
Tier A: Frontier AI labs
feasible
high confidence
Anthropic is the only observed actor and has the production weights.
Tier B: Indonesian AI labs
not feasible
high confidence
The work requires production weights that only the developer holds.
Tier C: Institution units such as ITB
not feasible
high confidence
The work requires production weights that only the developer holds.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
not feasible
high confidence
The work requires production weights that only the developer holds.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeClaude 3.5 Haiku with a cross-layer transcoder of about 30M features; hardware not given.
- DataProduction model weights, which only the developer has.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): not sourced in the notes.
- Time horizonMulti-year agenda.
- Shelf-life signalsSatisfying insight on only about a quarter of prompts tried
- Pipeline roleTier A produces methods and tools, some later open-sourced for small models (topic 16), which pass to Tier B and Tier C.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
On the Biology of a Large Language Model
Observed actors
Linked models
Links
R&D TopicMechanistic Interpretability, Frontier Scale
Use Casesnone recorded
Who Runs Itnone recorded
Papers
- Toy Models of SuperpositionNelson Elhage et al., arXiv preprint, 2022 founding
- Scaling and evaluating sparse autoencodersLeo Gao et al., arXiv preprint, 2024 state of the art
- Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language ModelsSamuel Marks et al., International Conference on Learning Representations, 2025, 2024 state of the art
- Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2Tom Lieberum et al., arXiv preprint, 2024 state of the art
- Mechanistic Interpretability for AI Safety -- A ReviewLeonard Bereska et al., arXiv preprint, 2024 survey
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- On the Biology of a Large Language Model transformer-circuits.pub https://transformer-circuits.pub/2025/attribution-graphs/biology.html
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Mechanistic Interpretability, Frontier Scale", https://aurion.semai.id/topics/mechanistic-interpretability-frontier-scale/, records of October 2026.