Interpretability on Small Open Models
Mechanistic interpretability methods applied to 1B to 9B open models.
Tier Grades
Tier A: Frontier AI labs
feasible
medium confidence
Anthropic and Google DeepMind are observed and released the tools.
Tier B: Indonesian AI labs
feasible
medium confidence
Tools are open-sourced for 1B to 2B models, though hardware needs are not stated in the notes.
Tier C: Institution units such as ITB
feasible
medium confidence
Open tools on 1B to 9B models suit academic work; hardware needs are not stated in the notes.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
stretch
low confidence
The notes give no anchor linking this to proof-of-concept or proof-of-value work, and SAE research was deprioritised by Google DeepMind.
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
- ComputeHardware requirements for circuit-tracer not stated; Gemma Scope SAEs on all layers of Gemma 2 2B and 9B, select layers of 27B.
- DataOpen weights of Gemma-2-2B, Llama-3.2-1B, Gemma 2 9B.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): not sourced in the notes.
- Time horizonWeeks to months.
- Shelf-life signalsGoogle DeepMind deprioritised fundamental SAE research in March 2025 after SAE probes underperformed linear probes out of distribution
- Pipeline roleTier A open-sources tools (circuit-tracer, Gemma Scope); Tier C and the open community apply them to small models.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
circuit-tracerGemma Scope
Observed actors
Linked models
Links
R&D TopicInterpretability on Small Open Models
Models
Gemma 2 2BGoogle, LLMGemma 2 9BGoogle, LLM
Llama 3.2 1BMeta, edge LLMGemma2-2b-ITGoogle, LLM+1 more
Gemma 2 2BGoogle, LLMGemma 2 9BGoogle, LLMLlama 3.2 1BMeta, edge LLMGemma2-2b-ITGoogle, LLM+1 moreUse Casesnone recorded
Who Runs Itnone recorded
Papers
- TinyStories: How Small Can Language Models Be and Still Speak Coherent English?Ronen Eldan et al., arXiv preprint, 2023 founding
- Sparse Autoencoders Find Highly Interpretable Features in Language ModelsHoagy Cunningham et al., arXiv preprint, 2023 founding
- OLMo: Accelerating the Science of Language ModelsDirk Groeneveld et al., arXiv preprint, 2024 state of the art
- Pythia: A Suite for Analyzing Large Language Models Across Training and ScalingStella Biderman et al., arXiv preprint, 2023 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- Open-sourcing circuit-tracing tools anthropic.com https://anthropic.com/research/open-source-circuit-tracing
- Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2 arxiv.org https://arxiv.org/abs/2408.05147
- deepmindsafetyresearch.medium.com/negative-results-for-sparse-autoencoders-on-d... https://deepmindsafetyresearch.medium.com/negative-results-for-sparse-autoencoders-on-downstream-tasks-and-deprioritising-sae-research-6cadcfc125b9
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Interpretability on Small Open Models", https://aurion.semai.id/topics/interpretability-small-open-models/, records of October 2026.