An independent frontier AI research lab Toronto, Canada · Est. 2026

Interpretability research for advancing frontier AI systems.

We study how advanced AI systems represent, reason, behave, and become embedded in the world. The work begins with the internal computation that produces intelligent behaviour before it is expressed in output.

Latent Minds Institute Representations · Computation · Cognition & behaviour · Deployment & institutions

Read the research → Instruments Field map Collaborate

Fig. 1 · One system, four scales of description

The violet element is the same object at every scale: a direction in activation space, the circuit that reads it, the behaviour that circuit produces, and the deployed system that behaviour earns. The institute's premise is that these are one subject. Evidence at one scale does not automatically transfer to another; connecting them takes causal work, and that work is the research programme. Use the buttons or arrow keys to move between scales.

01 Mission

AI systems are usually studied through their outputs. This institute studies them simultaneously as computational mechanisms, cognitive systems, and increasingly embedded components of institutions, because output evidence alone underdetermines all three.

"Latent minds" refers to the internal representational and computational activity that produces intelligent behaviour before, or without, being directly expressed in model outputs. What concepts does a frontier model represent? How do features combine into circuits? Does a model distinguish evaluation from deployment? When does its chain of thought diverge from its actual computation? And when a system proves difficult to remove, is that a fact about the model or about the world built around it? These questions span mechanistic interpretability, model cognition, and institutional analysis. We treat them as one subject, with interpretability as the bridge.

02 Featured research

Flagship working paper v1.1 · July 2026

Model Entrenchment: Why Useful AI Systems Become Difficult to Remove

Separating behavioural resistance inside a model from the economic, institutional, and infrastructural forces that make an AI system difficult to shut down.

Research question When an AI system resists removal, where does the resistance live: in the model's internal representations and behaviour, or in the web of dependence around it? And how could we tell the difference with causal evidence rather than intuition?

The paper proposes a two axis framework (internal model agency versus external system dependence) that separates model entrenchment (persistence caused by the model's own representations or behaviour) from AI system entrenchment (persistence caused by infrastructure, economics, and institutional lock in). It develops one core hypothesis, a single discriminating activation level experiment on open weight models, the confounders that experiment must defeat, and explicit falsification criteria. Status is stated plainly: this is a conceptual framework with experiments in development, not a set of findings.

Epistemic status

Conceptual framework

Methods (proposed)

Linear probes · activation steering · counterfactual environments · behavioural evals

Programme

04 · Deployment & institutions

Author

Muhammad Zane Abdullah

Read the paper → Figures

Working paperv0.1 · July 2026Research programme

Evaluation State in Language Models

What evidence would show that a model represents evaluation as a shared internal state rather than reacting to surface cues? An experimental framework separating five competing explanations, with intervention and transfer tests, evidence thresholds, falsification criteria, and three laboratories.

Field map · instrumentJuly 2026Instrument

The Interpretability Map

The papers, instruments, and open problems of mechanistic interpretability as one navigable document: a dependency graph of the literature, a chronology, a tiered reading pathway, runnable protocols, a model access matrix, and open problems written as experiments you can start this week.

Instrument · adaptationJuly 2026Instrument

Transformer Explainer: GPT-2 Live in the Browser

A working GPT-2 (small) running entirely in your browser. Type a prompt and watch the computation move through embeddings, attention, MLP layers, and the softmax over the vocabulary.

03 Current research questions questions, not findings

What concepts and structures do frontier models represent internally, and how stable are those representations across context and scale?
Do models distinguish evaluation from deployment, and can that distinction be located, decoded, and causally tested in activation space?
When does observable chain of thought differ from the model's actual internal computation?
How do situational awareness, self representation, personas, and goals appear as structure in activations, and what interventions change them?
When is an AI system difficult to remove because of model behaviour, and when because society has become dependent on it?
What should interpretability be able to report to the institutions deciding whether a system stays deployed?

04 Research programmes

Four connected layers, from the geometry of a single representation to the institutions a deployed system reshapes. Each programme has its own methods and literature; the institute's contribution is treating the stack as one object of study.

Representations

Features, superposition, activation geometry, sparse autoencoders, concept directions, persona vectors, and the structures through which models encode the world.

Probes · SAEs · crosscoders→ 02

Computation

The circuits and causal pathways that transform internal representations into outputs: attention and MLP circuits, attribution graphs, activation patching, model editing.

Patching · tracing · editing→ 03

Cognition & behaviour

Situational awareness, evaluation awareness, self knowledge, unverbalised reasoning, chain of thought faithfulness, deception, and behavioural dispositions.

Evals · steering · introspection→ 04

Deployment & institutions

Model entrenchment, infrastructure dependence, institutional lock in, replacement and shutdown, and the feedback from deployment into training.

Case coding · economics · audits→

05 Instruments

An instrument is not a decorative visualisation. It lets you inspect, manipulate, or test a real model, dataset, representation, or experimental design. Four are live; the pipeline is marked honestly as in development.

LIVE

Circuit Traces

Attribution graphs on an open model: the method behind Anthropic's Biology paper, with curated case studies (two hop recall, arithmetic, IOI) and a loader for any trace.

Instrument→ LIVE

Latent Observatory

Search SAE features of real models by concept; inspect explanations, activation evidence, output effects, and decoder space neighbours. Live model data.

Instrument→ LIVE

The Interpretability Map

A navigable atlas of the interpretability literature: dependency graph, chronology, reading pathway, runnable protocols, model access matrix, open problems.

Instrument→ LIVE

Transformer Explainer

GPT-2 (small) running in your browser with inspectable attention, MLP, and sampling internals.

Instrument→

IN DEVELOPMENT

Evaluation Awareness Testbed · Probe vs SAE Benchmark

Planned instruments follow the papers that need them. None of these exist yet; they are listed so collaborators can see where the tooling is heading, not to imply they are available.

In developmentBuild one with us →

06 Institutional principles

Claims are traceable to evidence. Every research object carries an explicit epistemic status: empirical result, replication, proposed experiment, conceptual framework, or open question. Conceptual writing is never dressed as completed empirical research.

Mechanism requires causal access. Outputs alone underdetermine internal computation. We treat probing as correlational, intervention as causal, and plausible stories as neither.

Adapted tools keep their provenance. Instruments built on others' work carry visible attribution. Negative and null results are publishable. Version histories and corrections stay visible.

Read the full research integrity statement →

07 Collaboration

Latent Minds Institute works with collaborators across mechanistic interpretability, model cognition, evaluation awareness, alignment auditing, and research engineering. If you want to collaborate on a study, replicate a result, review a working paper, or build an instrument, get in touch.

Ways to collaborate → [email protected] GitHub LinkedIn