The Interpretability Map · Latent Minds Institute

Part 1

The field as a graph

Six areas, foundations, features, circuits, frontier, auditing, and instruments, form a spine, and each paper, tool, or venue attaches to one of them. Cross links show where one line of work feeds another. The order runs roughly left to right: the residual stream gives the field its coordinate system; superposition explains why single neurons are hard to read; sparse autoencoders recover cleaner features; attribution graphs assemble those into computation; and auditing uses the whole stack to check what a model is actually doing.

Box 1 · Working vocabulary

residual stream: The shared vector space every layer reads from and writes to; the model's central communication channel.
superposition: Storing more concepts than dimensions as overlapping directions, the reason single neurons are polysemantic.
feature: A direction in activation space corresponding to a concept; recovered at scale by sparse autoencoders.
circuit: A causal pathway of features across layers implementing a computation; mapped with attribution graphs.
steering: Clamping a feature's activation to change behavior, the causal test that a feature is real.
activation patching: Swapping internal activations between runs to localize which components matter for a behavior.
transcoder / crosscoder: SAE relatives that model a sublayer's computation or span multiple layers; cross-layer transcoders are the substrate of attribution graphs.
linear probe: A supervised classifier trained on activations, cheap, strong baseline that any unsupervised method must beat.
alignment audit: A systematic investigation into whether a model pursues objectives its behavior conceals.

Where the field stands. Interpretability is now used inside real deployment work: Anthropic examined a production Claude model's internal features as part of its pre deployment safety assessment. The methods are still contested. DeepMind found that simple linear probes beat sparse autoencoders on some safety detection tasks and stepped back from foundational SAE research, while Anthropic's work moved forward through transcoders, crosscoders, and readouts that put activations into plain language. Both things are true at once: dictionary learning is useful and imperfect, and the frontier is largely an argument about which readouts to trust. The open problems in Part 4 come out of that argument.

Figure 1. The dependency graph of the literature (Field), the working loop of a practicing researcher (Protocols), and the open problems by difficulty and area (Problems). Filled circles are lobes; open circles are papers, instruments, and venues, sized by centrality. Select any node for its dossier below.

log as read

The striking property is acceleration: five years from manifesto to production scale features, then three major readout methods in the eighteen months.

Figure 2. Chronology of the primary literature, 2020–2026.

Part 2

The linear pathway

The literature is strictly cumulative, read the layers in sequence.

Part 3

The laboratory

The fastest way in is to copy a working notebook and start editing it. Four protocols cover the loop : tinker to build intuition, train a dictionary of features, test a method against ground truth, and read where the field publishes.

Protocol 1 · Tinker

Cache and inspect a model's activations

~30 min

The operation of the field: run a forward pass with hooks into every internal activation. Everything else, patching, probing, SAEs, builds on this cache.

# Colab, free GPU
pip install transformer_lens

from transformer_lens import HookedTransformer
model = HookedTransformer.from_pretrained("gpt2-small")
logits, cache = model.run_with_cache("The Eiffel Tower is in")
cache["blocks.7.attn.hook_pattern"].shape  # attention patterns, layer 7

Continue with the TransformerLens Main Demo notebook, then ARENA chapter 1. → TransformerLens · ARENA

Protocol 2 · Train

Load, then train, a sparse autoencoder

a weekend

Start by loading a pretrained SAE and inspecting real features; train your own only once you can interpret someone else's. Gemma Scope provides production grade dictionaries across every layer of Gemma 2.

pip install sae_lens

from sae_lens import SAE
sae = SAE.from_pretrained(              # pretrained dictionary
    release="gemma-scope-2b-pt-res-canonical",
    sae_id="layer_20/width_16k/canonical")
# then: encode activations, rank features by activation,
# and inspect top-activating dataset examples on Neuronpedia

API details move quickly, trust the library docs over any snippet, this one included. → SAELens · Gemma Scope · Neuronpedia

Reference · Access

Opening the hood: what you can actually experiment on

bookmark this

Interpretability requires internals. This is the honest access map, from full weights on your laptop to remote hooks on models you could never host.

Model / platform	What you get	Best for
GPT-2 small	full weights, every hook, runs on a free Colab GPU	first circuits, patching, the classics (IOI)
Gemma 2 / 3 + Gemma Scope	weights plus production-grade open SAEs at every layer (270M–27B)	feature work without training your own dictionaries
Pythia suite	16 models, identical data, 154 checkpoints each	how features and circuits form during training
OLMo	open weights and open training data	tracing a feature back to what the model saw
Llama-class via nnsight / NDIF	remote execution with full access to internals	frontier-scale questions without frontier-scale hardware
Neuronpedia	hosted features, attribution graphs, steering in the browser	zero-setup exploration and feature characterization
Goodfire Ember	hosted API for reading and steering features	steering experiments at scale, product prototypes
Claude / GPT / Gemini APIs	behavior only, no internals	behavioral baselines your mechanistic claims must beat

Protocol 3 · Test

Audit a model with a planted defect

1–2 weeks

The field's rare gift to newcomers: benchmarks with ground truth. AuditBench provides 56 models with implanted hidden behaviors; Anthropic's evaluation agent is open sourced. Run an existing method against a known defect, measure honestly, write up what failed.

# the shape of an auditing experiment
# 1. pick a target model with a known implanted behavior
# 2. probe: SAE features, activation patching, behavioral tests
# 3. score your hypothesis against the ground-truth defect
# 4. publish the negative results too, the field needs them

→ Automated auditing agents · Alignment Science blog · circuit-tracer

Protocol 4 · Read

Follow the field where it publishes

ongoing

Three venues carry nearly everything: Transformer Circuits for the canon, anthropic.com/research for the accessible summaries (read these first), and the Alignment Forum for work-in-progress and fast feedback. For mentorship into the field, MATS is the standard pipeline.

Part 4

The open problems, as experiments

Open Problems in Mechanistic Interpretability, written by researchers across many labs, and the 200 Concrete Open Problems sequence, which rates entry points by difficulty. Below are eight problems written as experiments you could start this week, each with a first set of steps. Filter by difficulty, and expand a card for the protocol.

¹ Calibrate expectations before committing months: the field's own leadership has moved from "fully reverse-engineer the model" to "partial understanding that genuinely helps evaluation, monitoring, and incident analysis." See the 80,000 Hours interview.

Appendix A

The index

Every bookmark in the atlas, tabulated by lobe. The same data drives Figure 1.

Citation

Citing this map

This is a living navigation document, not a peer-reviewed work. Cite the primary sources it points to. To reference the map itself:

@misc{interpretability-map-2026,
  title  = {The Interpretability Map},
  author = {Abdullah, Muhammad Zane},
  note   = {Interactive living document, Latent Minds Institute.
            Originally published under MO3 Research.},
  year   = {2026},
  month  = {July},
  url    = {https://latentmindsinstitute.com/instruments/interpretability-map/}
}