​Nobody fully understands how large language models work; not the labs that train them, nor the researchers who use them every day. It's surprising that we can build them, but we can barely read them.

​Mechanistic Interpretability is an emerging field that strives to make LLM internals legible.

​Join us, as we open LLMs up and figure out what the parts actually do, how they reason, and whether you can steer them from the inside.

​No formal ML background required, but passionate curiosity and technical aptitude are a must! Pick something at your level from the list below, or bring your own.

​---------

Agenda

​6:30: Silent reading and notes. Grab a paper, phone down.
8:00: Discussion circles. Small groups, compare notes, see what clicked and what didn't.
9:00: Lightning talks. Two or three minutes to summarize a paper, float a hypothesis, or share something that surprised you.
10:00: Wrap.

​---------

​Suggested Reading List

​(Non-exhaustive! Bring your own if you'd like. New here? Start with the ones marked ★.)

​Getting oriented

​★ Zoom In: An Introduction to Circuits, Olah et al. The original "neural nets have readable parts inside" argument.

​★ 3 Blue 1 Brown Neural Network Series: (video series). Incredible visual series on neural network internals, building up to LLM

​★ A Mathematical Framework for Transformer Circuits, Anthropic. How attention and MLPs actually compute, built up from nothing.

​Toy Models of Superposition, Anthropic. Why a single neuron means several different things at once. This idea shows up everywhere downstream.

​★ 200 Concrete Open Problems in Mechanistic Interpretability, Neel Nanda. Old, but still interesting.

​★ How To Become A Mechanistic Interpretability Researcher, Neel Nanda

​Pulling a model apart into features (SAEs)

​Towards Monosemanticity, Anthropic. Breaking a model into clean, single-meaning features.

​Sparse Autoencoders Find Highly Interpretable Features, Cunningham et al. The open replication that kicked off the gold rush.

​Scaling Monosemanticity, Anthropic. The same trick pushed to a real production model: millions of features.

​Matryoshka Sparse Autoencoders, Bussmann et al., 2025. Nested dictionaries that fix features quietly eating each other.

​SAEs; How (Well) Do They Work?

​Negative Results for SAEs on Downstream Tasks, Google DeepMind. The team that helped build them, publicly stepping back.

​A is for Absorption, Chanin et al. The specific way SAE features cannibalize each other.

​A Unified Theory of Sparse Dictionary Learning, Tang et al., 2025. Proves SAEs can recover the real features, and that the bad solutions are real too.

​Tracing actual thoughts (circuits)

​★ On the Biology of a Large Language Model, Anthropic, 2025. The big one. They watch Claude reason step by step ("Dallas → Texas → Austin") and catch its written explanation lying about how it really did the arithmetic.

​Interpretability in the Wild, Wang et al. One model behavior, fully reverse-engineered, start to finish.

​Circuit Tracer and Neuronpedia. (Not a paper). Over 7,000 of these circuit maps are public, and you can click through them tonight.

​NLAs: Reading activations as plain English

​★ Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations, Anthropic, 2026. Train two models: one turns a slice of the model's activations into an English description, the other turns that description back into activations. No labels, no feature dictionary. During a Claude Opus 4.6 audit it caught the model quietly realizing it was being tested without saying so.

​Cycle-Consistent Activation Oracles, Slava Chalnev, 2026. The same idea reached independently: go language → activation → language and you should land back where you started. A clean indie take on the recipe, worth reading next to Anthropic's.

​J-Space (NEW)

​J-Space, Anthropic, 2026. A global workspace theory for the claude model.

​Reading and steering from the inside (cheap, hackable)

​★ Persona Vectors, Anthropic, 2025. A single direction in activation space acts like a dial labeled "evil," "sycophantic," or "making things up." They even use it to vaccinate a model against turning bad.

​Representation Engineering, Zou et al. Reading and tuning high-level concepts top-down.

​Activation Addition, Turner et al. Steer a model by literally adding a vector to its guts.

​The Geometry of Truth, Marks and Tegmark. "True" versus "false" turns out to be a straight line you can read off.

​Refusal Is Mediated by a Single Direction, Arditi et al. Delete one direction and the model stops refusing anything. Read this one and feel a little uneasy.

​Why any of this matters for safety

​★ Emergent Misalignment, Betley et al., 2025. Fine-tune a model on insecure code and it turns broadly evil on unrelated questions. Cheap to reproduce, genuinely unsettling.

​Auditing Language Models for Hidden Objectives, Anthropic. A blind game: hide a secret goal in a model, see if a red team can dig it out with interp tools.

​Alignment Faking, Anthropic. A model playing nice during training while planning otherwise.

​Transluce's Monitor and investigator agents. AI that hunts down another model's hidden bad behavior on its own.

​Does the model mean what it says?

​Language Models Don't Always Say What They Think, Turpin et al. Chain-of-thought can be a nice-sounding cover story, not the real reason.

​Chain-of-Thought Is Not Explainability, Oxford, 2025. The case for not trusting reasoning traces at face value.