An observatory for internal computation
Latent Minds Institute is an independent frontier AI research lab. It studies what happens inside advanced models, and what happens to the world that deploys them.
The most consequential artifacts of this decade are systems whose internal workings their builders cannot fully describe. A frontier model's behaviour is the visible tip of an enormous computation: billions of parameters transforming representations across dozens of layers before a single token appears. "Latent minds" names that hidden activity, the representational and computational structure that produces intelligent behaviour before, or without, being expressed in output.
Studying it is not optional curiosity. Everything society wants to know about these systems, whether an evaluation measured what it claims, whether stated reasoning reflects actual reasoning, whether a system behaves differently when it believes it is being watched, whether it can be safely removed, is a question about internal computation, asked from the outside.
Outputs underdetermine mechanism. A model that writes "I don't want to be shut down" might be imitating its training data, playing a role the prompt implies, or expressing a representation of its own situation, the possibility worth taking seriously precisely because the others exist. The text is identical in all three cases. Behavioural evidence can establish that a pattern is systematic; it cannot establish what computes it.
That is why the institute's methodological spine is causal: probes are treated as correlational, interventions (patching, steering, ablation) as causal, and plausible stories as neither. Interpretability that stops at plausible stories inherits the failure mode of pre-registration-era psychology: compelling narratives, unfalsifiable claims. The field has the tools to do better, and the institute holds itself to them, including publishing the null results.
Model-level research that never touches deployment answers half the question. The systems being interpreted are simultaneously being woven into hospitals, banks, courts, and software supply chains. Whether a system can be removed is decided partly by its weights and partly by the world's dependence on it, and conflating those two produces bad safety arguments in both directions. The institute's fourth programme exists to keep the mechanistic work honest about where it matters: interpretability results should eventually be legible as evidence to the institutions deciding what stays deployed.
These commitments bind everything the institute publishes:
- Primary sources are preferred, and every claim must be traceable to evidence.
- Conceptual frameworks are labelled as conceptual frameworks. Proposed experiments are never presented as completed experiments.
- Replications identify every deviation from the original methodology.
- Adapted tools retain visible attribution to their original authors.
- Negative and null results are publishable and expected.
- Interactive explanations distinguish toy demonstrations from real model data.
- Version histories remain visible, and corrections are documented, not silently applied.
Every research object carries one of these status labels: Empirical result Replication Proposed experiment Conceptual framework Open question Interactive instrument
Muhammad Zane Abdullah, Founder. Toronto, Canada.
Latent Minds Institute is an independent research lab founded by Muhammad Zane Abdullah. The institute develops interactive research, conceptual frameworks, replications, and experimental tools for understanding advanced AI systems, and works with collaborators across interpretability, model cognition, and research engineering.
The institute is independent and self-funded. Research published before the institute's founding appeared under MO3 Research at mo3.ca; those objects carry their original publication dates and a provenance line, and old URLs redirect here.
Each research object carries its own BibTeX. To reference the institute generally: